Coexist, orchestrate, consolidate: adopting AI without rip-and-replace

Adoption plans that require a switchover fail for a boring reason: the switchover has to happen on a specific Tuesday, and there is never a Tuesday when a company can afford it.

The alternative is three phases, ordered so that each one is a stable resting place. You can stop at any of them and still be better off than when you started, which is what makes the sequence survivable.

Phase one — coexist

Your existing systems stay. The new layer reads them, and it does not write.

This is often dismissed as a trial, and it is not. Coexistence does one specific job that has to be done before anything else is safe: it builds the state model. Reading across your CRM, project system, payroll, and billing produces a picture that spans them, and the picture is the prerequisite for everything downstream.

It also produces the first honest surprise. Almost every company discovers in this phase that two of its systems disagree about something material — a customer count, an owner, a status — and has been operating on whichever one the person asking happened to open. That discovery on its own frequently justifies the phase.

What coexistence should not do is act. Read-only is not caution for its own sake; it is because you cannot govern actions against a state model you have not yet checked.

Connectors and what depth each one reads at are on integrations.

Phase two — orchestrate

Now the layer coordinates, and here is where the value concentrates.

Orchestration means the system coordinates your team, AI agents, workflows, and connected systems — with frontier models acting as workers rather than as the system itself. Work is compiled into missions, actions are taken through governed calls into the tools you already have, and the tools remain where they are.

Two properties make this phase safe rather than alarming.

Scoped access. Each agent gets only the context and tools its mission allows, rather than a blanket integration credential. The scope is the unit of governance, and it is what makes it possible to allow an agent into your CRM for one purpose without allowing it into your CRM.

Actions inside the real products. Agents work inside real products through governed tool calls, not by producing text that a person then retypes. The retyping step is where most AI pilots quietly lose their value: the output is good, the transcription is manual, and the manual step means the volume never grows.

This is also where external coding agents fit, if you have engineering. They can be governed over the same protocol: task capsules out, pull-request diffs back to a review desk, continuous-integration status registered as a check, and anything above a risk threshold gated on approval. The mechanism is on MCP.

Most companies should plan to live in phase two for a long time. It is not a waypoint.

Phase three — consolidate

Consolidation means retiring a system because the layer above it has genuinely absorbed its job.

This phase is real and it is oversold. The honest test is a single question, asked per system rather than in general: does anyone still open it?

If a tool is only ever written to by automation and read by nobody, it is a data store you are paying application prices for, and consolidating it is straightforward. If people still open it — because of a workflow, a habit, an integration outside your control, or a regulatory reason — then retiring it moves cost rather than removing it.

Consolidate the ones that pass. Leave the ones that do not. A company that ends up with six systems instead of eleven, all sharing one state model, has taken the win; insisting on one is a preference, not a strategy.

The two skipped-phase mistakes

Skipping coexist. Going straight to orchestration means governing actions against a state model nobody has audited. The failures this produces are the confident, well-formatted kind: correct actions taken against a wrong picture. They are also hard to diagnose, because the action log looks fine.

Skipping orchestrate. Going from reading straight to replacing means you never built the governance record that would justify the replacement. You are then making a large irreversible decision on the strength of a demo.

The order is not arbitrary. Each phase produces the evidence the next one needs.

How long each phase takes

Vendors give timelines. The honest answer is that each phase ends on a condition, not on a date, and the conditions are checkable.

Coexist ends when the state model reconciles — when the disagreements between systems have been found and either resolved or explicitly accepted with a note saying which source wins. For most companies this is weeks rather than months, and it takes longer the more systems there are, not the more people.

Orchestrate ends — if it ends at all — when a domain has enough action record to justify consolidating something. That is a function of volume, so it takes as long as it takes to accumulate actions. A low-volume domain may sit here permanently and should.

Consolidate is per system, not a phase you complete. Each retirement is its own small decision with its own answer to the does-anyone-open-it question.

The failure pattern is treating these as a schedule and moving on the date rather than on the condition. A phase-two rollout on a company still holding unreconciled state is the same mistake as skipping phase one, arriving three months later with more momentum behind it.

What each phase should produce

Phase one produces a state model and a list of disagreements between systems. If you finish it without a list of disagreements, you did not look hard enough.

Phase two produces an action record: what the system did, under what authority, what got rejected, what got escalated. This record is the input to every later decision about scope.

Phase three produces a smaller bill and a shorter list of places where truth lives.

The day-to-day of operating in phase two — the brief, the decisions, the missions in flight — is on run your business.