AI POC to Production Timeline: What the Gap Actually Costs

Your sponsor asks when it ships. You say "a few weeks" because the model works. Twelve weeks later the model still works and it still has not shipped, and you no longer have a credible answer.

The AI POC to production timeline is not one number. It is five gates, each with a different owner, a different duration, and a different failure mode. A pilot stalls at the handoffs between them, not inside them. This article gives you each gate, an honest range for how long it takes, who has to sign it, and the arithmetic for what a week of delay costs — so you can hand your sponsor a dated plan with names on it instead of a quarter.

Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, escalating costs, and unclear business value. Every one of those is a gate problem. None of them is a modelling problem.

The short answer

A well-scoped AI project takes 10 to 16 weeks from working pilot to production, assuming a dedicated team and no enterprise approval committees. That time is not spent improving the model. It is spent clearing five gates: a data contract, an evaluation baseline, an integration contract, on-call ownership, and a rollback path. Each gate has a named owner outside the AI team, which is exactly why pilots stall — the work sits in someone else's queue and nobody has agreed what "done" means.

Why pilots stall at handoffs, not at the model

The demo works. That is the problem. A working demo convinces everyone that the remaining work is small, so nobody schedules it.

What actually remains is a series of transfers of responsibility. The AI team hands data requirements to a data owner. The data owner hands access to platform engineering. Platform hands a deployment to whoever will carry the pager. Each transfer needs someone to accept an obligation they did not ask for, and each one queues behind that person's existing commitments.

The distinguishing feature of a project that ships is not technical. It is that every gate below has a name against it before week one, and the owner has agreed to the definition of done. Projects that fail can usually name the model. They cannot name the person who accepted the on-call rotation.

The five gates

Gate 1: The data contract

Done means: a written agreement on the fields the system consumes, their types, their null behaviour, their freshness guarantee, and who to contact when the schema changes. Not a sample extract. An agreement.

Owner: the data owner for the source system, not the AI team.

Typical duration: 1 to 3 weeks. Longer where the source is a vendor system whose schema changes without notice.

What it blocks: everything. A model trained against an undocumented extract will drift the first time an upstream team adds a column, and you will discover it in production.

The trap is accepting a one-off extract to keep momentum. That extract is a snapshot of a system that is still moving. Without a contract you have not acquired data, you have borrowed some. This is the same discipline covered in our enterprise AI data strategy work, and it is the gate teams most often skip.

Gate 2: The evaluation baseline

Done means: a held-out set with agreed ground truth, a measured human baseline on the same set, and a written threshold that decides go or no-go. The threshold is agreed before the number is known.

Owner: the business owner who will be accountable for the outcome.

Typical duration: 1 to 2 weeks, plus whatever labelling takes.

What it blocks: the go-live decision, and every argument after it.

Without a pre-agreed threshold, launch becomes a negotiation about whether the number is good enough, held at the moment when everyone is invested and tired. Teams that set the bar first ship faster because there is nothing left to debate. Teams that set it afterwards discover that "good enough" moves whenever someone senior looks at an error.

Measure the human baseline even when it is uncomfortable. Most teams find the human number is worse than assumed, which changes the conversation from "is the AI accurate" to "is it better than what we do now."

Gate 3: The integration contract

Done means: documented APIs at both ends, agreed authentication with a credential rotation plan, defined rate limits, and a stated latency budget the receiving system will honour.

Owner: platform or the owning team for each target system.

Typical duration: 2 to 5 weeks, and this is where estimates break down most often. Integration is routinely 40% or more of total effort.

What it blocks: any end-to-end test that means anything.

The specific failures are unglamorous and predictable: tokens that expire on a schedule nobody documented, rate limits discovered under load, source formats that change quietly, and latency spikes that only appear at real volume. Build handling for each of these deliberately. Our integration patterns guide covers the shapes that survive contact with production.

Gate 4: On-call ownership

Done means: a named rotation, a runbook for the three most likely failures, an escalation path, and an agreed severity definition for a model that is up but wrong.

Owner: the engineering manager whose team will carry the pager.

Typical duration: 1 to 2 weeks to negotiate, and it is a negotiation.

What it blocks: go-live, in any organisation with a functioning production discipline.

This is the gate teams forget entirely, and the one most likely to add a month. Asking a team to take on-call for a system they did not build, cannot debug, and did not budget for is a real request. It needs to be made early, in writing, with the runbook attached. Made late, it becomes a blocking objection from someone with every right to raise it.

Note the severity question is genuinely hard. A model that returns a confident wrong answer is not "down," so it will not page anyone unless you have defined what wrong looks like and instrumented it.

Gate 5: The rollback path

Done means: a tested route back to the previous process, a documented trigger for using it, and a named person authorised to pull it without convening a meeting.

Owner: whoever owns the business process, not the AI team.

Typical duration: 1 week, if the old process still exists.

What it blocks: the staged rollout, and your ability to launch calmly.

Rollback is what makes a staged rollout safe rather than theatrical. Shadow mode, then 5%, then 25%, then full traffic only works if reversing is cheap at every step. If the manual process was decommissioned during the pilot — which happens more than people admit — rollback is not a runbook entry, it is a rebuild, and your risk profile is entirely different from what you told your sponsor.

The gate table

Gates 1 and 2 can run in parallel. Gate 3 usually cannot start until Gate 1 closes. Gates 4 and 5 can run alongside Gate 3 but need to start early because they are negotiations, not tasks.

That is where 10 to 16 weeks comes from, and why "the model is done" tells your sponsor almost nothing.

What a week of delay actually costs

Say "it will take another few weeks" often enough and it stops registering. Put a number on it and the conversation changes, because delay is a real cost that nobody is currently booking.

Cost of one week of delay has two parts:

Burn. The fully loaded weekly cost of everyone waiting. A team of four at a fully loaded rate of $3,000 per week each is $12,000 per week, whether or not the gate moves.

Deferred benefit. The weekly value the system would have produced. If the business case is $600,000 of annual benefit, that is roughly $11,500 per week not being captured.

For that example, a week of delay costs about $23,500, and a gate that slips a month costs roughly $94,000. Against those numbers, three days of a data owner's time to write a schema contract is not overhead. It is the cheapest thing on the plan.

Two things to be careful about. Deferred benefit is only real if the benefit is real, so use the number from the approved business case rather than a fresh optimistic one. And burn only counts people who are genuinely blocked; if half the team moves to other work during the wait, halve it. The point is not the precision. It is that the number exists at all, because an unpriced delay always loses to a priced cost elsewhere.

Turning gates into a dated plan

Your sponsor does not want a quarter. They want a date and a name. Build it in this order:

1. List the five gates and put a name against each. If you cannot name the owner, that is the first thing to fix, not a detail to resolve later.
2. Get each owner to agree the definition of done in writing. One paragraph each. This is the step that converts a plan into commitments.
3. Ask each owner for their own estimate, then add their queue time rather than their working time. A gate that takes three days of effort but sits two weeks in a queue is a two-week gate.
4. Sequence with the dependencies above, run what can run in parallel, and start the negotiations (gates 4 and 5) first because they have the longest human latency.
5. Publish the cost of delay per week next to the plan so slippage has a visible price.
6. Review weekly against gates, not percentages. "Gate 3 of 5 closed, gate 4 blocked on a named person" is a status report. "70% done" is not.

What makes the estimate wrong

The 10 to 16 week range assumes a dedicated team, documented APIs at the integration targets, data access without a committee, and stable scope. Each broken assumption adds time in a predictable way:

• Enterprise data approval committee: add 3 to 6 weeks, mostly queue.
• Undocumented or vendor-controlled integration target: add 2 to 4 weeks.
• Part-time team: multiply everything, do not add. A half-allocated team takes more than twice as long because context reloading is not free.
• First implementation of this kind: add 50% to 100%. This is a learning cost and it is real.
• Regulated process: add validation and evidence time, and pull the compliance owner in at gate 2 rather than gate 5.

If several apply, say so early. A sponsor told about a committee in week one hears a constraint. Told in week nine, they hear an excuse. For a view on whether to build in-house or work with a partner, the learning-cost multiplier is usually the deciding factor.

Practical next steps

1. Audit your stalled projects by gate. For each, name which of the five gates is open and who owns it. Most teams discover their pilots are stacked on the same two gates.
2. Get owners in writing before week one. The gate list is a staffing request, not a checklist.
3. Price the delay and publish it. Put it in the same document as the plan.

If your pilot is stuck, it is almost certainly not the model. Read why AI POCs fail for the patterns underneath, and why AI projects fail for the organisational versions.

FAQ

How long does AI POC to production typically take?

Ten to sixteen weeks for a well-scoped project with a dedicated team and no approval committees. That time is dominated by five gates rather than by modelling: a data contract (1-3 weeks), an evaluation baseline (1-2 weeks plus labelling), an integration contract (2-5 weeks), on-call ownership (1-2 weeks of negotiation), and a rollback path (about a week). A first implementation of a given kind typically takes 50% to 100% longer, and a part-time team takes more than twice as long rather than proportionally longer.

What actually causes POC-to-production delays?

Handoffs, not technical difficulty. Every gate is owned by someone outside the AI team — a data owner, a business owner, a platform team, an engineering manager who will carry the pager — and the work queues behind their existing commitments. Integration alone routinely consumes 40% or more of total effort. The gate teams most often forget entirely is on-call ownership, because asking another team to support a system they did not build is a negotiation that takes weeks if it starts late.

What does a week of delay cost?

Add the fully loaded weekly burn of everyone blocked to the weekly value the system would have produced. A blocked team of four at $3,000 each per week is $12,000, and a $600,000 annual business case defers roughly $11,500 a week, so a week costs about $23,500 and a month about $94,000. Use the approved business case figure rather than a fresh estimate, and only count people genuinely blocked. The precision matters less than having the number, because an unpriced delay always loses to a priced cost somewhere else.

What is a data contract and why does it gate everything?

A written agreement covering the fields consumed, their types and null behaviour, a freshness guarantee, and a named contact for schema changes. It gates everything because a model built against a one-off extract is built against a moving system. When an upstream team adds or renames a column, an extract-based pipeline fails silently in production while a contract-based one fails loudly in testing, or does not fail at all because the change was communicated.

Get from POC to production

Applied AI Studio specializes in production deployments. We have moved dozens of AI projects from demo to shipped product by closing these gates in order. If your project is stuck, let's diagnose the blockers.