AI implementation engagement model: the first 90 days
An embedded AI engineer engagement runs in six stages: a short call, an on-site Tech Audit, a transformation roadmap with success criteria in writing, a kickoff, about three months with the engineer in your team three days a week, then handover to one of your engineers. "Done" means a system in production, in use, that your team runs alone.
Most of an embedded AI engineer engagement is decided before the engineer arrives. The use case, the baseline and what "done" means are settled in advance, on your site and in writing. Below is the whole path, from the first call to handover: who does what at each stage, what your company provides, and how you know it is finished.
What does the engagement look like, from the first call to handover?
| Stage | How long | What you provide | What it ends with |
|---|---|---|---|
| 1. The [contact form](/contact) and a short call | A form, then a short call | One process and one person who knows it | A custom offer: AI consulting or a Tech Audit |
| 2. [Tech Audit](/services/tech-audit) | One morning on site, report in 3 working days, one-hour handoff | Four hours with the people who do the work, a view of systems and data | 3–5 ranked opportunities and what not to build |
| 3. [Transformation roadmap](/blog/ai-transformation-roadmap) (Deep Dive) | Two visits on site, then a presentation to leadership | Real data extracts, the sponsor and leadership at the presentation | A written roadmap for each use case and a decision on the first one |
| 4. Kickoff | One meeting | The sponsor, your paired engineer, the access list | The milestone, the adoption metric and the on-site days, in writing |
| 5. The engineer in your team | About three days a week, three months | Access, the process owners, test users | The use case in production, and in use |
| 6. Handover | The last weeks of the three months | Your engineer, who takes the system over | Code, documentation, evaluation set and one of your people who knows them |
What happens before the engineer arrives?
Three steps, each ending with a decision that is yours. You can stop after any of them.
- The contact form and a short call. You tell us about the process in the contact form; after a short call we send you an offer that fits: AI consulting or a Tech Audit.
- The Tech Audit. Four hours on your site with the people who do the work. We see the processes, systems and data as they are. Within three working days you get a short report: 3–5 ranked opportunities, what each would take, and what is not worth building.
- The transformation roadmap (Deep Dive). Two visits on site, research on your sector and a written roadmap for each use case: the preconditions, the AI Act classification, the baseline, the production milestone and the adoption metric. We present it to your leadership in one hour. This is where success criteria get written down, before anything is built.
The order is not ours. Colin Jarvis, who leads OpenAI's forward deployed engineering team, describes the same three phases: scoping, with "A couple of days onsite"; validation, with criteria and evaluations set in advance; then delivery on site "typically for a few days per week" (The Pragmatic Engineer). We have sized them for a mid-sized company.
What gets written down at kickoff?
Before the engineer's first day, we sign one page together: the engagement charter. Without it we do not start.
- Your sponsor: who unblocks access and decisions.
- The production milestone: which process, which users, by which date.
- The adoption metric and its baseline, measured in week one.
- Your paired engineer, and how much of their time they spend alongside ours.
- The access list, who approves each access, and who handles escalations.
- The on-site days: week one, go-live and handover, plus the days agreed with you.
What do the first 90 days look like, week by week?
| Weeks | The Sapio engineer | Vlad Tudor, engagement lead | Your team |
|---|---|---|---|
| 1 (on site) | Access and environment, a written plan on day one, the baseline measured, the first real cases in the evaluation set | Kickoff with the sponsor, confirms the charter | The sponsor, the paired engineer, the process owner for a few hours |
| 2–4 | A first version of the workflow runs in your environment around day 10; the evaluation set grows; a written update every week | A 30-minute check-in with the sponsor every two weeks | The process owner answers questions and checks outputs; your engineer pairs |
| 5–8 | Integration with the real systems, logging, monitoring, cost per request, a rollback plan; first users on real cases | The monthly steering review; in week 8, the conversation about what comes next | A small group of users tests on real work |
| 9–10 (go-live on site) | Go-live for the users named in the charter | On site for go-live, follows adoption | Users trained; the sponsor clears blockers |
| 11–13 | Watches where people route around the system and fixes the reasons; prepares the handover package | Checks the milestone and the metric against the charter | Your engineer takes the system over at handover |
The weekly rhythm is deliberate. Sierra, which deploys customer-service agents, writes the same about the period after launch: "We continue to meet weekly with customer stakeholders" (Sierra). The difference between a demo and a system in use is made in the weeks after go-live, not before.
What does your company provide?
- A sponsor with authority, available for half an hour every two weeks and for the monthly review.
- One of your engineers, paired from week one. They take the system over at the end.
- The person who knows the process, a few hours a week, for questions and for checking outputs.
- Access to the systems and to real data in week one, not week four.
- A few users who test on their real work before go-live.
- For personal data or regulated systems, your data protection officer involved from the start.
What does "done" mean?
"Done" is written into the charter on day one, not negotiated at the end. The engagement is finished when all four hold:
- the named process runs in production for the named users, by the agreed date;
- the adoption metric is met against the baseline measured in week one;
- you hold the code, the documentation, the evaluation set, the runbooks and the cost model;
- you have a written review: what shipped, and what the next step would take.
A system that works but is not used is not "done". That is why so few projects pay back: only 25% of AI initiatives have delivered the expected ROI, and only 16% have scaled enterprise-wide (IBM, 2025 CEO study).
What happens if something goes wrong?
We tell you early and in writing. Each of these signals triggers a conversation with the sponsor in the same week:
- no production access after week two;
- no agreed adoption metric by week two;
- the engineer drifting into tickets unrelated to the use case;
- a change of sponsor, or a security review blocking the work for weeks.
There are three options: reset the milestone, pause, or end the engagement. An engineer who has drifted into closing tickets is staff augmentation under our name, so we would rather stop.
What happens after the 90 days?
Three options, all your call: we continue monthly on the same use case, we move to the next use case on the roadmap, or we stop and your engineer runs the system. The conversation happens in week eight, so you never decide under pressure on the last day.
What the engineer does, and why they work as a pair with an engagement lead, is explained in what is an embedded AI engineer. Every step, from the contact form to the engineer in your team, is on our AI consulting page. If you want to know whether your process is worth this path, tell us about it in the contact form; after a short call we send you an offer that fits: AI consulting or a Tech Audit.
Sources
- The Pragmatic Engineer, "Forward deployed engineers", 12 August 2025
- Sierra, "Shipping and scaling AI agents", 25 July 2024
- IBM, 2025 CEO study with Oxford Economics, 6 May 2025
Only 25% of AI initiatives have delivered the expected ROI, and only 16% have scaled enterprise-wide (IBM Institute for Business Value, 2,000 CEOs, published 6 May 2025).
Frequently asked questions
How long does an embedded AI engineer engagement take?
The steps before it, the short call, the Tech Audit and the transformation roadmap, usually take a few weeks. The engineer then works in your team about three days a week for three months, on one use case. After that you continue monthly, move to the next use case, or stop.
What does the client company need to provide?
A sponsor with authority, an internal engineer who pairs from week one, the person who knows the process for a few hours a week, access to the systems and to real data in week one, and a few users who test before go-live.
What does it mean for the engagement to be done?
The named process runs in production for the named users by the agreed date, the adoption metric is met against the baseline, and you hold the code, the documentation and the evaluation set.
Do we have to go through every step?
The Tech Audit, yes: we do not place an engineer without having seen the process on site. The transformation roadmap is for companies weighing several use cases. When there is one clear use case, its success criteria go straight into the kickoff charter.
What happens if the engagement stalls?
The signals are written down in advance: no production access after week two, no agreed adoption metric, or an engineer drifting into tickets unrelated to the use case. Any of them triggers a conversation with the sponsor in the same week: reset the milestone, pause, or stop.
Want to discuss a project?
Book a free discovery call with the Sapio team.