Forward Deployed Engineer
Owns
The system works in production: architecture, code, integration, performance, your security review.
Lives in
Your repos, your CI/CD and your on-call reality.
Fails if
It does not run reliably on your infrastructure.
Forward Deployed AI Engineering
Senior engineers inside your environment. Production AI shipped there.
Working software in 30 days. Production in 90. Measured on your P&L — and priced on the outcome, not the hours.
Get your AI readiness score →2 minutes. Then book a 90-minute scoping session — no cost, no commitment.
Does this sound familiar?
The pilot demoed beautifully in March. It is still not in production.
The integrator’s meter keeps running. The outcome date keeps slipping.
The platform you licensed is still waiting for the engineering to connect it.
If you are nodding, the next five minutes explain why this keeps happening — and the fastest credible way out.
The failure rate is not a talent problem. It is a delivery-model problem.
>80%
of AI projects fail — twice the failure rate of non-AI IT projects.
~30%
of GenAI projects were forecast to be abandoned after proof of concept by the end of 2025.
The five in a hundred that report P&L impact.
95%
of enterprise GenAI pilots show no P&L impact, self-reported.
Each square is one project in a hundred.
Over the same window, the models got dramatically better. So whatever kills these projects, it is not model quality.
The demo and production are different sports.
Clean samples. Dirty data from systems nobody has cleaned in a decade. Nothing to integrate. Legacy ERP and the internal systems around it. No permissions. Row-level permissions and your IAM. No review. Security review, vendor review, change control. No SLA. Uptime, on-call and someone to page at 03:00.
The demo
Production
01Clean samples.
Dirty data from systems nobody has cleaned in a decade.
02Nothing to integrate.
Legacy ERP and the internal systems around it.
03No permissions.
Row-level permissions and your IAM.
04No review.
Security review, vendor review, change control.
05No SLA.
Uptime, on-call and someone to page at 03:00.
Five constraints the pilot dodged. Every one of them is an engineering problem inside your environment, and none of them can be solved from outside it.
Nobody owns the last mile. That is the whole story.
The model
One build. Working software in 30 days. Production in 90.
We onboard in your stackHardenedDay 010306090First thin slice liveProduction cutover
- Day 1–10We onboard in your stack: access, repos, a map of your real data.
- Day 30First thin slice live — real users, real data.
- Day 30–60Hardened: permissions, audit, security review.
- Day 90Production cutover: runbooks and handover. Yours to extend.
- 01Everything lives in your repos from the first commit.
- 02Weekly demos on working software — never a status deck.
- 03Handover is not a phase at the end; it is how we work from day one.
The market already converged on this model.
Cited as evidence the model works — these are not Datacean clients.
- Palantir
1,300+ bootcamps in 2024 — five days, the customer’s own data, a working use case by Friday. Two decades of forward deployed delivery, now its core sales motion.
- OpenAI
A dedicated forward deployed engineering unit. They hold the best models on earth and still concluded they need engineers physically inside customer environments to reach production.
- C3.ai · Scale AI
The same delivery model — engineers deployed into the customer’s environment, paired with the customer’s own team, across heavy industry, defense and finance.
The pattern won for a reason. Production AI is an in-environment engineering problem, and only embedded engineers sit next to the constraints.
You have three conventional options. None of them owns the last mile.
- 01
Hire your own team.
Upside
Permanent capability.
The catch
6–12 months to hire in the hardest talent market in tech — the pilot backlog does not wait.
- 02
Buy a SaaS platform.
Upside
Fast start.
The catch
It ends at the API. Integration, data quality and permissions — the actual gap — stay on your side.
- 03
Big-4 or systems integrator.
Upside
Scale plus cover.
The catch
Billed on effort: senior people sell, junior people build, and the meter runs either way.
Five phases. Every gate is a decision point — not a renewal.
- 01
Strategy Sprint
3 weeks
Hard exit gate
A ranked portfolio and a go/no-go — or we stop.
- 02
Diagnostic Pilot
4–6 weeks
Hard exit gate
The agreed metric hit on your real data.
- 03
FDE Build
8–16 weeks
Hard exit gate
Live in production, through your change controls.
- 04
Value Realization
4–8 weeks
Hard exit gate
Value signed off by your sponsor, not by us.
- 05
Enablement
2–4 weeks
Hard exit gate
Your team runs a full cycle without us.
0481216 weeks
Stop at any of the five gates and keep everything — code, docs, the trained team.
The system has to work in production — and be worth building.
A Forward Deployed Engineer owns the first; a Deployment Strategist owns the second. Two jobs, deliberately not one person.
Deployment Strategist
Owns
The system is worth building: use-case selection, the success metric, stakeholder alignment, adoption.
Lives in
Your steering meetings and your business units.
Fails if
It runs perfectly and nobody uses it.
Build work and stakeholder work compete for the same hours, and the person who wrote the system should not be the one grading its business value.
Security and governance
Built to pass your vendor review, not to dodge it.
- Data boundary
Everything stays inside your perimeter. Nothing is copied to Datacean systems, including for our own development.
- Identity
Named individual accounts in your IdP, least-privilege and time-bound. You revoke unilaterally, at any time.
- Model hosting
Your decision: open weights in your perimeter, your cloud’s in-tenancy endpoints, or an API under your contract. No external inference if your policy says no.
- Endpoints
Your devices or your VDI, under your EDR and DLP. Nothing is stored, cached or processed on Datacean hardware. MFA through your IdP.
- Personnel
Named Datacean employees only, background-screened to your standard, under individual NDAs. No subcontracting without prior written consent. You approve every individual and can require replacement.
- Any stack
AWS, Azure, GCP, on-premise, air-gapped. In air-gapped environments, engineers work on your hardware and code arrives through your media-ingress and review process.
- Incidents
Any suspected incident involving Datacean personnel or access is reported to your security team within 24 hours of detection. We operate under your incident response process.
- Offboarding
All accounts, keys and tokens revoked within 24 hours of exit sign-off, a joint access-log review, and a written attestation that no data, credentials or repositories persist on any Datacean-controlled system.
A DPA is signed before access, and a BAA where PHI is in scope. Practices are mapped to SOC 2 and ISO 27001 control sets and evidenced control by control in your review.
At the end, you own everything. Unambiguously.
What stays with you
- 01Source code, in your repositories from the first commit.
- 02Pipelines and infrastructure as code.
- 03Runbooks and operational documentation.
- 04The architecture rationale — the why, not just the what.
- 05Prompts, configurations and evaluation suites.
Full IP assignment. You can run, modify and extend all of it without us: no runtime licence, no per-seat fees, no phone-home. Any pre-existing Datacean tooling is declared in the SOW before use, under a perpetual, royalty-free, transferable licence.
Every engagement is built to end
- 01The SOW commits your team to handling 90% or more of operations and changes without us by month 12, measured quarterly and reported to your sponsor.
- 02We certify two or three of your people as internal trainers, so new staff never needs a vendor invoice.
- 03Then it is your call: full handover, which is the default, a light advisory retainer, or the next use case through the same gates.
A vendor you can leave cheaply is a vendor you can trust to stay for the right reasons.
Pricing
We don’t sell hours.
The day-rate model is hours multiplied by rate multiplied by months. The meter runs whether it works or not, you buy effort, and 100% of the delivery risk stays with you. We sell a measured result instead. Fixed fees stay fixed and overruns are ours. The variable part is tied to a number signed before kickoff: a baseline measured jointly, a target agreed before we start, a frozen method, a fixed window, a named arbiter, always capped.
| Strategy Sprint | Three weeks. Use-case selection, feasibility on your data, a build plan. | Fixed fee. | $45k – $70k |
|---|---|---|---|
| Production Pilot | Ninety days. One use case, from measured pilot to first production deployment. | Fixed, milestone-gated. | $120k – $250k |
| Embedded FDE | Senior engineers embedded in your teams, your cloud and your repositories. | Annual. | $400k – $1.2M/yr |
| Value Share | Only where a hard savings metric exists. Arbiter-verified, first 12 months post go-live. | 10–20% of verified savings, capped. | Cap set per SOW |
| Enablement | Your engineers trained and certified to run and extend the system. | Annual retainer. | $60k – $120k/yr |
Indicative — every engagement is scoped per client, in writing.
Where the risk sits
- 01Miss the pilot gate and the final 30% milestone is not invoiced. We fund up to four more weeks of remediation at our own cost, and that 30% is invoiced only if the frozen-method measurement then passes.
- 02If savings fall short, the value share is a percentage of arbiter-verified savings at whatever level they land, up to the cap. Zero verified savings, zero share. No cliff, no threshold.
- 03Either way you keep everything: code, pipelines, documentation, your repositories, your cloud, your IP from day one. There is nothing to hand back, and the embedded agreement exits on notice.
The pilot price quotes the annual rate in advance, so nothing is renegotiated from a position of dependency, and 50% of the pilot fee is credited on conversion within 30 days.
Who delivers
The principal, not the pyramid.
[FOUNDER: photograph]
Miguel Fierro spent ~10 years at Microsoft working on AI as a Forward Deployed Engineering Manager — participating in over 100 projects deploying AI workloads inside customer environments and generating over $500M of business impact. He is the creator of Recommenders, the most popular open-source recommendation library on GitHub. He has made over 200 interviews for AI profiles. He holds a PhD in Robotics (UC3M with King’s College London, best doctoral thesis award) and executive education from MIT Sloan. He delivers personally.
The person who sells the work is the person who shows up.
- Forward Deployed Engineering Manager, Microsoft
- 100+ projects deploying AI inside customer environments
- $500M+ of business impact
- Creator of Recommenders
- 200+ interviews for AI profiles
- PhD in Robotics, UC3M with King’s College London, best doctoral thesis award
- MIT Sloan executive education
Honest terms
When this works — and when it doesn’t.
Succeeds when
- 01A named owner with decision authority meets us weekly.
- 02Access lands in week one, because the approvals were done during contracting.
- 03Scope targets one measurable outcome, not a platform vision.
- 04Your engineers get real hours to pair with ours.
Dies when
- 01The sponsor delegates to a committee.
- 02Access requests queue with no executive push behind them.
- 03Success is undefined, so everything becomes in scope.
- 04The code lands but nobody on your side ever touched it.
A named owner, a weekly decision cadence, and 20–40% of two of your engineers. If we cannot get these, we tell you before you spend money. We have declined work on these grounds. We put this on the page because vendors never do.
One next step: the sprint scoping session.
90 minutes, your sponsor plus one technical lead. You leave with a written scope, a fixed fee and a start date. No cost, no commitment; the sprint starts within 30 days of a signed scope.
Get your AI readiness score →2 minutes, then pick a time.