An AI Transition Field Guide

Six questions an operator has to answer.

A field guide, not a playbook.

A playbook implies a complete and exhaustive method, and that is not our claim. Our work with clients has taught us that different businesses face different exposure depending on target customers, workflow, contract structure, data rights, margins, and competitive position, and no single sequence of plays works across the board with all of those dynamics to bear.

What this guide intends to do instead is help you recognize the terrain, diagnose your own position, read the signals, weigh the legitimate choices, judge how much time you have, and decide how to respond. Our hope is that you find it actionable without pretending to be exhaustive.

We have organized it around six questions we believe operators actually have to answer, in the order a leader would ask them. The four plays we refer to throughout (re-architect the business model, rebuild the product for agentic readiness, realign go-to-market, and restructure the organization to run two engines) are high-conviction actions arising from our work and analysis.

What is actually changing?

On February 3rd, 2026, roughly $285 billion in market value left software, financial-services, and asset-management stocks in a single day (Bloomberg). US software fell 6%, its largest one-day drop since the April tariff selloff, on the release of a single AI tool, not on any deterioration in anyone's reported numbers.

In our opinion, the underlying impact matters more than the headline. The worst damage was not to per-seat SaaS. LegalZoom fell 20%, Thomson Reuters 16%, LSEG 13%, FactSet 10.5%, Morningstar 9%, while Salesforce and ServiceNow, the canonical seat-priced names, fell about 7%. The market did not sort on "is this priced per seat." It sorted on "can the model replace the need for people to do this work."

The selloff was set off by Anthropic releasing a legal-automation tool, so the professional-services and legal-data names took the sharpest hit, the companies directly in that product's path. But the contagion reached financial data, asset management, and horizontal SaaS the legal tool did not touch. That this spread beyond the immediate target is what marks this as a capability sort rather than a same-sector reflex.

For two decades, software helped a human do the work (especially knowledge work). What is emerging is software that actually does it. We call the second thing Software-as-a-Worker, or SaaW, and the distinction is not cosmetic: it changes what you sell, who you sell it to, and what a unit of value is.

Every play we write about is a move from the left column toward the right. Pay attention to the last cell in the last row: the labor budget is an order of magnitude larger than the software budget SaaS companies compete for now, and that is the strategic prize in the transition.

The evidence will show in the labor market before the financials

AI's effect on seat-based software is already visible in one place and absent in another. The work that seats used to pay for is being automated, and the headcount doing that work is falling now. The revenue attached to those seats is not, at least not yet, because enterprise contracts delay the transmission. From the market analysts' lens, the bears expect revenue to be falling already, and it is not, while the bulls read healthy revenue as safety when the mechanism that will erode it is already running underneath all of this. Three facts lay it out.

Fact 1: Seat-heavy headcount is falling. US customer-service-representative employment fell 130,180 jobs, or 4.8%, between May 2024 and May 2025 (BLS Occupational Employment and Wage Statistics). Secretaries and administrative assistants fell 1.8%; wholesale and manufacturing sales reps 2.3%. Across the eighteen most AI-exposed occupations, employment declined while total US employment grew 0.8%. Strip out medical secretaries, whose numbers rise on healthcare demand, and the other seventeen fell for the second consecutive year. BLS projections from July 2026 extend it: CSRs down a further 5.5% by 2034, with AI adoption named as a driver.

Occupation-level decline is not proof of AI causation: offshoring, macro softness, the post-2021 over-hiring correction, and reclassification move these same numbers. Three things separate the AI reading from those alternatives: the decline concentrates in exactly the occupations most exposed to current model capability and existing agent automation, not across the board; it runs against a rising total-employment backdrop, so it is not macro; and the two-year duration fits a mechanism that builds as capability ramps, where a one-year move would read as noise. What none of this establishes on its own is the coupling to revenue, which is why we treat it as a leading signal rather than a closed case.

Fact 2: Seat-based revenue is not falling. Through H1 2026, HubSpot's NRR was 105%, up 1.5 points year over year with customers up 16%. monday.com's NDR was 110% with double-digit seat growth. Atlassian's NRR was above 120%. Workday's gross retention was 97%. Asked directly about seat compression, HubSpot's management said customers bought more seats across every hub through the year, and monday.com reported no degradation in demand relative to seats.

Fact 3: Multiples repriced, not the revenue. Workday beat on the quarter in February 2026 and fell 10% to a fifty-two-week low. Two different things get called "seat changes," and only one is actually happening. Seats are being repriced: vendors shifting from per-seat billing to usage, action, and outcome models: real, happening now, and our read from speaking with sales leaders is that they are doing it to earn more per customer, not because seats are shrinking. Seats are not yet compressing. Net retention is flat to rising, and executives deny compression on the record. The falling headcount is the early signal of the second, not the first.

Four tangible shifts to note

Underneath the headline, the SaaS-to-SaaW move shows up as four specific shifts. This is a proprietary framework, synthesized from our client work against current BCG, Battery Ventures, and Bessemer reporting. Each is a classic SaaS advantage that turns into a liability unless you deliberately re-found it.

• Accessibility to autonomy. The browser was the SaaS distribution breakthrough; the agent is its AI-era equivalent. You now create the conditions for agents to work, not just for people to log in.
• Per-seat to value-metric pricing. Recurring revenue averaged across many customers, easily priced by public markets, now works against you. When one agent carries the load of several seats (which would logically lead to revenue lost), value has to eventually attach to work delivered rather than to human logins.
• Self-service UI to agent-ready ecosystem. Interfaces and feature checklists are being commoditized by code generation. The emerging moat is robust APIs, structured data, and the interconnections that let AI agents, yours and your customers', read, write, and execute directly. Agents will still need good UI to command and orchestrate but this UI will be different from the UI previously used by a person to do the work.
• Best practice to hyper-personalization. Standardized workflows give way to context-aware software that adapts per customer, per role, per moment, learned from usage.

Previous SaaS advantages were built around the human operator. SaaW advantages that take these four shifts as leverage are built around the work itself. The plays in Section 3 are the recommended response; each individual play is customized, so read on.

Where am I exposed, and where do I have an advantage?

Most writing on this transition is generically bearish on incumbents, which is both wrong and useless. The incumbent's problem is not that it lacks assets. It is that every asset has a liability welded to it. Getting specific about which is which is the difference between a winning strategy and a panic.

The five Janus realities

So you do not hold five assets with emerging liabilities. Each thing is more like a Janus; a face that looks forward and looks back. So you hold five Janus realities, each with an asset face and a liability face.

The revenue base. Asset: cash flow, and the ability to fund a multi-year transition from operations rather than dilution. Liability: it comes from seat-based licensing, which is exactly what Play 1 exists to dismantle. Every point of seat revenue you convert is a point of transition funding you spend.

The go-to-market machine. Asset: distribution, renewal motions, procurement relationships, and executive access a startup spends years trying to gain. Liability: a sales force compensated to sell seats, calling on departmental budget holders, with a comp plan that will fight the new model harder than any competitor.

The product and data estate. Asset: proprietary data, ranking above features and workflow integration in Brief 1's moat hierarchy. Liability: the architecture holding it was built for human-mediated workflow and dashboards, even old legislation; and you may not have the contractual right to use it the way the flywheel requires (see play 1's permission problem).

Trust, and the industry credibility you have earned. Asset: security reviews passed, compliance standing, an operating discipline that ships. Liability: an org chart, a comp structure, and often a board, all optimized for predictable SaaS metrics, exactly the review process that can starve anything new.

The bench, your human capital. Asset: people who hold the domain knowledge and operating history together, the one pairing an AI-native challenger cannot hire for at any price. Liability: this is the only asset on the list that can walk out the door. The other four are held in place by contracts, architecture, or inertia. This one is held in place by a soft labor market, which is a condition, not a moat, and it can change at any time.

The diagnostic: seven questions to help you understand your exposure

Score each with: 1 (weak), 2 (partial), or 3 (strong). Twenty-one points available.

1. 01Pricing exposure. What share of revenue depends on per-seat licensing, and what happens to NRR if a customer cuts relevant headcount 40%? Score 1 if over 70% is seat-linked with no variable component; 3 if under 30%, or you already bill on work delivered.
2. 02Agentic readiness. Can a third-party agent discover and execute your core functions via API today, without a human touching your UI? Score 1 if your API is a reporting API; 3 if an external agent completes your top three jobs-to-be-done end to end.
3. 03Data position. Do you hold the ground truth of the work, the record of what was decided, by whom, on what input, with what result? Score 1 if telemetry only; 3 if you hold the decision record and can attribute results to it.
4. 04Data rights. Do your contracts, DPAs, and BAAs permit cross-tenant training and the retention windows a feedback loop needs, and what share of ARR would need renegotiating data rights? Score 1 if you do not know; 3 if you have the rights or have renegotiated.
5. 05Outcome accountability. If you had to charge only for successful outcomes tomorrow, would customers accept your definition of one, and would your contracts survive it? Score 1 if you have no defensible definition of an outcome; 3 if you have one in a signed contract.
6. 06Organizational insulation. Does a team exist that could ship an AI-native product without being cannibalized by the core roadmap's quarterly demands? Score 1 if no such team exists or it reports into the core P&L; 3 if it reports to the CEO on its own KPI set.
7. 07Talent exposure. If hiring conditions improved sharply next quarter, who leaves, and what walks out with them? Have you asked, and do you know what your strongest people are optimizing for? Score 1 if you rely on the market to hold them and have not asked; 3 if you have a declared talent strategy, a named critical population, and a published route into the new work.

Four results and their positions

Exposed (7 to 10). High seat concentration, reporting-grade API, telemetry rather than ground truth, no insulated team. Your revenue is more fragile than your financials show. Start with the dual engine (Play 4): you have no vehicle, and everything else will be absorbed by the core roadmap within two quarters.

Instrumented (11 to 14). You have real data and some programmatic execution, but your pricing and your org are unchanged. This is the most common position in mid-2026, and the most deceptive, because your assets look strong so the urgency reads as low. Stand up the engine first (Play 4), then price into it (Play 1). Resist the temptation to start with the product rebuild (Play 2): it is where engineering-led teams instinctively go, and the play where the most money buys the least structural change.

Dual-Engine (15 to 18). A working part of the company already exists to take advantage of this moment, as a separate unit with its own metrics or with one variable pricing component live. Run the pricing and product plays in parallel inside that forward AI unit, with the guardrails from Play 4. Your failure mode is not slowness, it is erosion: your AI engineers get pulled back to help the core, or the AI unit shows up in the same dashboards as the legacy business and gets starved.

AI-Native (19 to 21). Work-metered or outcome-priced revenue can be material, agents execute core jobs without a human, the feedback loop runs. The focus is to realign go-to-market: the product market fit works, the sales motion just needs to catch up.

What choices do I need to make differently?

There is no ONE answer to this transition. While we focus on the plays in this guide, we know there are fundamentally four legitimate routes to take, and a board should evaluate them as a set before committing to any plays at all. As a leader of a SaaS company, choosing your primary route is the first decision, before executing on the plays below. A route is the path to the plays and there are four of them. Once you settle on the route, the plays make sense. In this brief we will focus on the plays for the first route.

The four routes you need to decide on first

The routes are not mutually exclusive; partner-while-you-build is common. But the choice has a clock on it, because Harvest and Buy both price off a seat-revenue multiple that is compressing. The option you have today is not the option you will have in eight quarters, and that is why route choice cannot wait for certainty.

The plays

When do I need to act?

Most transition advice gets timing wrong in both directions. The people telling you to reprice everything this year are reacting to a stock multiple shifts; they're likely panicing and too early. The people telling you your numbers are fine are reading a lagging indicator; they don't hear the train coming until its too late. Neither is watching the actual clocks, and there is more than one.

Four clocks, running at different speeds

Naming four instead of one gives you permission to sequence you response. The revenue clock gives you room on repricing; you do not have to touch the whole book this year. The talent clock gives you no room at all on the org and the bench, you should rethink org design and motivation very soon. That asymmetry (more time on pricing, less time on people) is the entire argument for doing things in an order rather than all at once.

The revenue clock: how to estimate how much time you have

The lag between headcount moving and revenue moving is not mysterious. It is contractual: seat reductions reach your revenue only at renewal, so the delay is roughly the average remaining term on your book. We use a simple model with clients to turn that intuition into a number they can plan against. The annual drag on gross retention, once it begins, is approximately:

Worked through: at S = 40%, h = 5%, T = 3, r = 33%, a customer arrives at renewal with about 14% fewer relevant heads, and the drag is roughly 1.9 points of gross retention in the first affected year, rising as more of the book renews into the same reality. Against 10 to 15 points of expansion, that is invisible for a while, and then it is not. We use h = 5% here because it is close to the customer-service decline the BLS data already shows. Substitute your own.

Be clear about what this model does and does not do. It assumes the headcount decline holds steady across the term, that renewals land evenly rather than in lumps, and that expansion elsewhere does not mask the drag. Real books violate all three to some degree. Treat it as a lens for sizing the window, not a promise about any single quarter. Run across the contract terms common in enterprise and SMB SaaS, roughly one to three years, and it lands in that 18-to-36-month range before a seat-heavy book has substantially repriced.

How much does the engine cost?

Before the sequence, the question every CFO asks first: what does standing up Engine 2 (Invent and Disrupt) actually cost. No incumbent discloses a clean operating budget for a ring-fenced AI unit, so we build the number from cost drivers that are measurable.

The team dominates the cost. The pattern for a new AI unit inside an incumbent is small, autonomous, and CEO-adjacent, closer to a skunk team than a division. A credible minimum viable Engine 2 is roughly ten to eighteen people (depending on the size of our company): a unit lead reporting to the CEO, four to seven applied-AI and agent engineers, one or two ML engineers, one or two PMs fluent in models, agents and evals, a designer, a forward-deployed engineer who sits with early customers, and legal and data engineering embedded rather than borrowed. Smaller is a pilot, not an engine. Much larger rebuilds the bureaucracy the ring-fence was meant to escape.

Senior AI and ML talent moves everything else. Fully loaded, including equity, benefits, and overhead, budget roughly $300,000 to $500,000 per senior technical person, higher in the Bay Area and New York (Carta 2025 startup-compensation data and BLS employer-cost loading). Inference and compute is the second line, and it behaves unlike classic SaaS: per-token prices are falling fast, but agentic workloads consume many times the tokens of a chatbot query, so total spend rises as the product scales. Model inference runs roughly a fifth to a quarter of AI product cost at scaling-stage firms (ICONIQ, 2026). Put together, a ten-to-eighteen-person engine runs on the order of $6M to $18M fully loaded in year one; a leaner eight-to-twelve-person launch team, $3M to $8M.

We recommend framing the commitment as a share rather than an absolute: roughly five to ten percent of R&D, or one to three percent of operating cash flow, for the ring-fenced unit. That sits below the share of R&D companies now point at AI generally, because Engine 2 is a new-growth bet rather than the AI features the core ships, and it echoes the old three-horizons guardrail that reserves about a tenth of resources for new-growth options. Large enough to matter, small enough that the board can carry it through a soft quarter without a special conversation. Whichever structure you choose, stage it against milestones rather than the calendar: gate year one on a shipped agentic capability and first design-partner pilots, gate year two on first paid deployments and a per-query gross margin trending the right way.

The sequence

When we analyzed the plays, we numbered them in the order that builds an investment arguement: model, product, go-to-market, organization. Working with clients, the order for doing turned out to be different. Two things kept surfacing as the real constraints, not the pricing or the architecture: the talent risk is the largest and the fastest-moving, and organizational change is far harder than any of the technical work. For that reason we now usually recommend standing up the organization first. Play 4 leads.

So the order is: stand up the engine (the talent structure that produces it), price into it, build into it, then sell what it produces. Section 2 tells you where to enter. Each phase carries a gate: the thing that has to be true before the next phase is worth starting.

Phase 0, this quarter, whatever your position. Run the diagnostic and publish the score (CEO). Audit data rights across the ARR base (General Counsel). Model seat exposure at 20% and 40% headcount decline using the arithmetic above (CFO). Inventory the ground-truth data you hold (CPO/CTO). Gate: the exec team agrees on a position, and the transmission window is a number you can take to the board.

Phase 1, quarters 1 to 2. Stand up the engine (Play 4). Charter and fund Engine 2: mandate, reporting line, P&L boundary, its own KPI set, a named team, and a run rate sized against the share target above. Ring-fence talent in writing with a named escalation exception. Embed legal from day one. Publish the talent contract from Section 6 alongside the charter. Gate: the board has approved the charter and the distinct metrics, the money is committed for four quarters, and the talent contract has been said out loud to the whole company.

Phase 2, quarters 2 to 4. Price into it (Play 1). Choose the work metric before the outcome metric, with the cost-to-serve model behind it (CPO/CFO). Model worst-case inference cost and set the exposure cap (CFO). Launch platform fee plus work metric in Engine 2, run the hybrid bridge in Engine 1. Gate: a contract is signed on the new structure, and you can produce a bill the customer accepts without manual reconciliation.

Phase 3, quarters 2 to 4, in parallel. Build into it (Play 2). Audit agent-executability of your top ten core functions and expose two end to end (CTO). Rebuild data infrastructure for streaming, retrieval, and feedback. Ship one agent-native capability to production, not to demo. Renegotiate data rights on the Phase 0 accounts. Gate: an external agent completes two core jobs without a human, and one capability is live with real customers and a measured success rate.

Phase 4, quarters 3 to 4 and beyond. Sell it (Play 3). Move messaging from feature utility to work execution (CMO). Identify executive budget owners and re-align sales incentives (CRO/CFO). Build the proof apparatus (CPO/CTO). Gate: the comp plan pays on work-metered revenue without a penalty, and a customer-facing outcome report exists.

What should I measure to know whether I'm ahead or behind?

There is a saying that a goal without a plan is just a wish. The same is true of a destination without a way to measure progress toward it. Most SaaS boards are still measuring the last era: DAU, feature adoption, seat count. Those tell you how the business you are leaving is doing, not how fast the one you are building is arriving.

Two layers matter. The macro dashboard from Brief 1 still runs in the background: outcome-pricing adoption, production deployment rates, regulatory movement on agent liability, and model-cost trajectory. Those tell you where the market is. But the signals that tell you where you sit inside your own business, and there are six of them. Review them quarterly.

The six internal signals

The first four measure how fast Engine 2 is growing.

1. 01Executable core. Share of your core jobs-to-be-done that an agent can execute without a human touching the UI. Target: above zero and rising every quarter.
2. 02Work-metered revenue as a share of total. Target: 2 to 5% and growing.
3. 03Outcome-priced revenue as a share of total. Target: above zero in at least one segment. Tracked separately from work-metered revenue on purpose, because merged, the number can be hit with metered API pricing that changes nothing structural.
4. 04Training-usable data captured per month under permissive terms. Target: rising month over month.

The last two measure how fast the assets funding all of it are depleting.

1. 05Latent attrition in the critical population. Share who have shipped agentic work in the last two quarters, internal applications per Engine 2 posting, and the three-offers answer (Section 6). Target: measured twice a year at minimum, with a trigger level set in advance.
2. 06Engine 1 erosion. Seat NRR, seat count, and seat-heavy headcount at your top 20 accounts. Target: tracked, with a trigger level set in advance.

Four signals tell you how fast the new engine is coming up; two tell you how fast the fuel is running out. Between them, they set your true timeline. A program that watches only its own progress cannot tell the difference between having time and being out of it.

What would prove us wrong

We hold a strong point of view in this guide, and a point of view that cannot be disconfirmed is not worth much. So here is what we are watching that would tell us we are wrong. If you are testing our thesis against your own business, these are the same things you should watch.

• Seat revenue holds through 2027 while seat-heavy headcount keeps falling. That would mean enterprises are redeploying freed capacity rather than cutting licenses, and the revenue coupling mechanism is weaker than we think. Our response: slow Play 1, keep Plays 2 and 4, which pay off under every scenario.
• Agent-actions become the dominant new metric, and stay that way. If by 2028 outcome pricing is still confined to support deflection, Play 1's destination is wrong even though its direction is right. Our response: stop at the metered stage. It is a complete and valid business model, not just a waypoint.
• Inference costs fall far and fast enough that COGS stops being a constraint. The margin caution in Play 1 goes away. Our response: accelerate. This is the pleasant version of being wrong.
• Ring-fenced units underperform absorbed AI programs. The play we watch hardest. But ring-fenced units almost never fail because separation was wrong; they fail because it was half-done: judged on the core's clock, cut off from the core's data, weak on talent, or asked for revenue too early. Our response: rule those out first (see Appendix A). If they are genuinely clean and the unit still lags, a single engine is the honest answer.
• Your best people stay through a hiring recovery even without the talent contract. Then the soft market was not the only thing holding them, and the full contract was insurance you did not need. Our response: keep the two pieces that serve Engine 2 regardless (the rotation that moves domain knowledge in, and the real Engine 1 mandate), and drop the rest.
• Your own top-20 accounts show no seat-heavy headcount decline over four quarters. Every other test leans on national data; this one runs on your own customers, the clearest read you will get. Our response: if your customers' headcount is genuinely flat, you have more time than the thesis implies: slow Play 1, keep Plays 2 and 4, and re-run this test every quarter.

If none of your signals ever move against you, it is very likely you are measuring the wrong things.

What do we carry forward?

Most transformation advice tells executives to abandon the old business and run at the new one. Our experience points the other way. The assets that make a successful transition possible are the ones you already own, sitting inside the business you are being told to leave behind. The job is not to escape the old model. It is to take apart its liabilities without breaking the assets attached to them. So the last question is not what to change, it is what to protect while you change everything else. Two things have to hold.

The assets that fund the change have to survive it

Go back to the trap in Section 2: the cash flow paying for this transition comes from the pricing model the transition is meant to dismantle. That is not a clever contradiction. It is a practical constraint, and it sets the pace. You convert each liability while the matching asset is still paying for the work, and never faster than that asset can carry. Reprice the book before you can measure the work, or re-architect before you know which capabilities agents actually call, and you have spent the asset without converting the liability. The five realities from Section 2 are not just a diagnostic score. They are the list of things that still have to be standing when you reach the other side.

Of the five, the one that moves fastest, and the one you cannot buy back, is your data. Every quarter your customers' work runs through your system instead of someone else's agent is a quarter you hold the moat. That is the competitive clock from Section 4, and it is the reason the product rebuild in Play 2 earns its keep under every scenario, including the ones where the pricing thesis turns out wrong.

The bench: the asset that can resign

Four of the five realities are held in place by contracts, architecture, or inertia. Your bench, your people, are held in place by a soft labor market, which is a condition that can change. It is the only asset on the list that can walk out, and it will walk out faster than the data migrates. The ring-fencing rule from Play 4 creates a problem it does not solve: ring-fencing staffs Engine 2 and manufactures a second problem: capable people left in Engine 1, reading the org chart accurately, concluding that the interesting work has moved somewhere they were not invited. Run the structure without a talent strategy beside it, and you will staff Engine 2 by quietly demoralizing the engine that funds it.

No single study measures the ring-fenced-unit talent drain directly; the mechanism is a synthesis of our experience and three literatures that converge on it. Organizational ambidexterity is the structural half: separating an exploratory unit from the core works only when the senior team runs on common-fate incentives anchored on the whole firm rather than the new unit's wins (O'Reilly and Tushman, 2008). Relative deprivation is the human half: in a study of a merged company, employees' sense of relative disadvantage predicted intent to leave, fully mediated by weakened identification with the organization (Cho, Lee and Kim, 2014). And Christensen's account of why incumbents fail is, underneath, a talent-allocation story: the best people follow the projects that enhance their careers (Christensen, 1997).

The leavers. Your strongest people going to AI-native companies outright. You will not out-compensate a challenger's equity, and you should stop trying. What you have that they do not is the thing your data moat is made of: real domain problems, real customers, real consequences, and a decade of proprietary data to build against. Say that specifically, in the room, or you are competing on the one axis where you lose.

The stayers under duress. Good people who want new problems, are frustrated to be on the core, and are staying because the market is soft, with full knowledge that they will move when it turns. This population does not show up in your regretted-attrition rate, because the soft market is suppressing the departures, not the intent. Gallup put the share of US employees watching or actively seeking at 51% as of May 2024, the highest since 2015, even as actual voluntary turnover had stabilized. This is not a low attrition rate. It is a deferred one, and the deferral ends on a clock you do not control.

The talent contract: five commitments

Declare the critical portions of these commitments to the whole company whenever possible.

1. 01A published route into Engine 2. Selection by tap on the shoulder produces one motivated team and a resentful company. Publish the criteria, run it on a cadence, and build a return path, so a rotation into Engine 2 is not a one-way career bet. Doing this makes it a motivational asset not a liability.
2. 02A real Engine 1 mandate, staffed like one. Engine 1's AI remit is rebuild work, not maintenance: agentic tooling on the core, internal automation, technical-debt reduction at a pace the old staffing never permitted. If that mandate is real, fund it and say so. If it is not, do not pretend; your engineers can tell inside a sprint.
3. 03A skills guarantee, with time and money attached. What your good people are optimizing for is not equity. It is not being obsolete in three years. Hours in the week, a training budget, and production work to apply it to are cheaper than replacing them and more credible than a retention bonus.
4. 04The truth about the roles that shrink. Some functions on both sides will be smaller in three years. People leave faster when they suspect a plan they have not been told about than when they are told a hard one. Name what you know, name what you do not, and commit to a date when you will say more.
5. 05A named critical population. A retention strategy that treats everyone as critical protects nobody. The intersection that matters is people who hold deep domain knowledge and can build agentic systems, the pairing an AI-native challenger cannot hire for and most wants. Identify these people by name, know what each is optimizing for, and review the list quarterly at executive level.

The ring-fence of the engine 2 team and a rotation plan are not in conflict. Play 4 bars Engine 2 engineers from being pulled back into Engine 1's escalations; commitment 1 builds a rotation between the engines. The purpose of the ring-fence is to block unplanned, reactive pulls: the 2 a.m. escalation, the legacy feature request that jumps the queue. The rotation is planned, scheduled, and chartered: a defined tour with a return path, staffed in advance. The rule is "no unplanned movement out of Engine 2," and it belongs in the charter in those words.

Grounded in primary sources.