Book a working session
Forward deployed engineers (FDEs) work with your team to build and deliver real business outcomes with the power of AI.
Here is what six of your teams could be saying once AI agents are doing real work for them, from refunds settled at 3 a.m. to the oldest code changed every week.
People who see more enterprise AI than almost anyone said why. Here are their own words. And when an AI agent goes wrong, it goes wrong in public.
"Are there examples of organizations who have been able to demonstrate production level systems and do those developments show a return in lower cost, efficiency or better top line?"
"How will we discover, govern and stop agents if need be. Perhaps the biggest security concern ATM."
The questions customers asked him, after "north of 200 meetings in Europe with customers and technologists".
"Decisions, processes, and workflows often live in people's heads and tribal knowledge that need to get turned into unstructured data for agents."
"IT teams are finding increasing success embedding full engineers into the business functions (essentially internal FDE) that go and implement agents into the internal workflows."
On "effectively the #1 problem for AI agents in the enterprise", then after a dinner with IT leaders of large enterprises.
AI agents now act on real systems. When one goes wrong, it's in public.
Put it live, and show the return.
Write down what your experts know.
Make it safe.
Here's everything it takes to get real business outcomes.
Week 7One stream waters every terrace. Each harvest is weighed against today’s number.
Point at a part and its things light up on the farm. Press a name under Who and that person waves. Press a stop on the line to move the season; the bead shows where it is. On the farm: press a gate, a trial bed, the wind pump, the irrigator, a crow, a tree, the bell, the hive, the hen; run the pointer through the field; point at the scale and each crate is traced back to its row; drag a cloud to make it rain. Any number is an example.
An FDE team of 5 to 7, inside your company until it’s live. Together they cover Ground, Orchestrate and Prove, AWS’s three FDE pathways.
We partner with domain experts from your industry, who draft the rules your experts then check.
On your side: who signs the scorecard, who check the answers, and who pair with ours from week 1 and run it after hypercare.
We build the loop, the tests and the harness. AI agents get a lot done.
By day people decide. Your people and ours review the work and write down what’s next.
At night AI agents do the long hours in a safe copy, with only the tools they need, against a number to beat. Only what beats today reaches your team, and you approve what ships.
AI agents earn more freedom only with evidence, and lose it on a serious slip. How we work, not a measured result
Our method for building, checking and running AI work, set up in your own AWS account from week 1.
Every piece of work goes through the same jobs and checkpoints.
The jobs build the harness: everything around the AI model, from its tools and memory to its policy and evals.
It stays with you, so the next outcome starts from here.
Parts we’ve already built, so the time goes on your outcomes.
Axiom gives every AI agent one meaning of your data:
your business terms in one catalogue;
each person’s access enforced by the data layer rather than the prompt;
every number traced back to its query.
Contact centres also get the Connect accelerators.
One page your owner signs by week 4: the outcome, the measure, today’s baseline and the target.
It’s the go-or-stop decision, and what our fee is measured against after that.
Your team runs it, or we do. AK Run looks after incidents, cost and drift, and keeps improving it. Its Agent ControlPlane keeps every AI agent behind one gate: approved models only, every call traced to a person, cost by team.
Weeks 1–3 · Agree
We sit with the people who do the work and write down how they decide: the rules, the exceptions and what must never happen.
Answers "live in people's heads" · Aaron Levie
Week 4: your owner signs the scorecard. Go, or stop and keep everything we found.
Weeks 2–6 · Ground
Every AI agent gets one meaning of your data, through Axiom, and sees only what it's allowed to.
Answers "live in people's heads" · Aaron Levie
Weeks 4–8 · Orchestrate
Sometimes the simplest thing is plain software or one model call, not an AI agent. Our FDEs direct the AI agents doing the heavy lifting, and check their work.
Answers "production level systems" · Nikesh Arora
Week 8: running on your real data, in a test copy of your systems. Your owner says go.
Weeks 8–12 · Control and Prove
It's tested many times over on cases no AI agent has seen. Your experts check it, and we attack it before anyone else can. Every AI agent can be found, limited and stopped.
Answers "discover, govern and stop agents" · Nikesh Arora
Week 12: ready to go live. Your people accept the measured results and practise stopping it.
Week 13 on · Run and Improve
It goes live to a small group first, then to everyone. Then hypercare: we stay on call beside your engineers until they run it on their own. Every failure it meets becomes a new test, and the learning stays with you.
Answers "show a return" · Nikesh Arora
By week 4 your owner signs a scorecard: the outcome, how it's measured, today's baseline, the target and what counts as a failure. If the numbers aren't there by then, we stop, and you keep what we found. From then on, the scorecard decides what we're paid. That's Guaranteed AI Outcomes.
ScorecardIllustrative
Customers get the right answer first time, with the right price for their contract.
MeasureConversations resolved with no correction, on a weekly sample your team leads check.
Day120
53%resolved with no correction
Our feeThe target is met.
Drag the coin to any day. Press the sign at week 4 to see a stop.
What a CFO, a CTO, a head of data and a CEO ask us first. Got another? Bring it to the working session.
Your CFOWhat does it cost, and how big is the team?
Armakuni An FDE team of 5 to 7 for about 13 weeks, then hypercare. The price is set per outcome, once the scorecard is agreed.
Your head of serviceCan you rescue the pilot we already have?
Armakuni Often, yes, and it's a good place to start. In weeks 1 to 3 we check it against a scorecard: what to keep, what to fix, and what's missing before it can go live. If it isn't worth finishing, we say so at week 4.
Your CTOWho are the engineers?
Armakuni Our forward deployed engineers, based in London, Florida and Ahmedabad. The lead and at least one engineer are with you every week. The lead owns the outcome; the others connect your data, build, prove and hand over. Each one leads a team of AI agents and checks their work.
Your CTOWhich AWS services, and whose account?
Armakuni Yours, from week 1. We usually build on Amazon Bedrock and Amazon Bedrock AgentCore. Amazon DataZone and AWS Lake Formation hold what your data means and who may see it. Every AI agent gets its own identity in Amazon Bedrock AgentCore, with least-privilege AWS IAM, and policy at the gateway checks every tool call.
Your CTODoes it work if we're not all-in on AWS?
Armakuni We build on AWS first, by design: our accelerators run there. The method works on other clouds too, but the accelerators don't yet, and we'll say so up front. Where your data lives in Databricks, we work with it there.
Your CISOOur auditors will ask. What do we show them?
Armakuni Evidence the system wrote as it ran. AWS Lake Formation enforces who may see which rows, not the prompt. Policy at the gateway checks every tool call an AI agent makes. Each release records what was approved: the instructions, the model, the tools and the policies. The kill switch and the rollback are rehearsed. The evidence pack maps each control to the rules you answer to: GDPR, HIPAA, SOC 2 and the EU's DORA. Mapped, not certified: your auditor decides.
Your CFOWhat happens to the fee if you miss?
Armakuni The part of our fee tied to the outcome is paid only when the scorecard shows the result. Before that, week 4 is a real exit, and nothing we found is lost.
Your head of dataWhat do we keep, and what do you keep?
Armakuni Everything we build stays in your own account, with full source: the code, the AI agents and their harness (the MCP tool servers, the policies, the tests and the eval sets), the domain model, the scorecard and the runbooks. At handover our access is removed. We keep only the generic patterns we learned, such as test patterns and attack suites. Never your data or your business logic.
Your CEOWhen shouldn't we use you?
Armakuni When one model call or a fixed workflow does the job. We'll say so and stop there. The same if nobody owns the outcome, or the data can't be reached in about four weeks. Better to hear it in the first conversation than in week 10.
Press "Do this for us" on anything above, or pick one here, or type your own idea. This is what a first conversation with us covers.
Bring a pilot that's stuck, or an outcome you want. We'll tell you plainly whether it's worth doing.
Two pages: every job the work goes through, what we check at each, and the two checkpoints before AI agents act and before it goes live. Version 2.5.