KLM Consulting builds supervised AI agents for energy and industrial operations — and the systems architecture they have to live inside. We take the work ourselves; nothing is brokered out.
We get called in after the pilot. The model is usually fine. What failed was everything around it — and it fails in a small number of recognisable ways.
Knowing which one you are in changes what the next eight weeks should cost.
The demonstration ran on a clean extract. Production runs on thirty years of scanned tickets, four naming conventions, a mainframe export and a spreadsheet one person maintains. Nothing was ever built to reach across them, so the agent answers confidently from a fraction of the truth.
Accurate enough to be interesting, not accurate enough for a scheduler, a measurement analyst or an HSE lead to sign. With no defined approval step and no named approver, there is no route from interesting to operational, and the pilot renews forever.
Reading is the solved part. Writing an approved change back into the scheduling system, the ERP or the ticket queue is an integration project with its own permissions, failure modes and rollback — and it was never in the budget. So a person retypes what the agent found.
An agent is a small amount of reasoning wrapped in a large amount of plumbing. We build the plumbing first, because that is where the risk and the cost actually sit.
Each part below is something you can inspect, test and switch off independently.
Autonomy is not a maturity score to climb as fast as possible. It is a decision about what a mistake costs, made once, in writing, before anything is switched on.
The agent reads across your records and answers questions, with the source page attached to every claim. It changes nothing. This is where a well file search, a drawing lookup or a contract question belongs — and it is where almost every engagement should begin, because it earns trust on read-only ground.
The agent prepares the action — the reschedule, the reconciliation, the permit check — and hands it to a named approver with its reasoning and its evidence. Edge cases route to a human. They are never resolved quietly. Most of the value we have delivered sits at this level.
The agent executes within limits you wrote: which actions, which value ranges, which counterparties, which hours. Anything outside the fence stops and escalates. Everything it did is reversible and logged. We recommend this only for narrow, high-frequency work where the fence is genuinely enumerable.
These are jobs with a clear trigger, a bounded set of records to consult, and a person who already signs off today. That combination is what makes an agent worth building; its absence is what makes one a science project.
Watch nominations against berth windows, tank availability and prior line-ups. Surface the clash while it is still cheap, propose the reshuffle, and put it in front of the scheduler with the conflict shown rather than asserted.
Reconcile meter tickets against recorded movements, flag unaccounted volume the day it appears instead of at month end, and attach the gauge reports and tickets that produced the variance.
Ask across decades of scans in plain language. Every answer returns the page it came from. Read-only, which is why it is usually the first thing we build.
Pull commercial terms out of a bid pack, mark what changed between revisions, and write the summary into the system of record rather than a mailbox.
Read the request, verify contractor certifications and the isolation list, route to the authorised signatory. Issuance stays with a person, always.
Match, flag the exceptions, and leave the clean ones alone.
Draft the handover from the day's records; the outgoing lead edits and signs.
Notice the pattern across purchase orders before the renewal conversation, not after.
Everyone in this market promises guardrails. Almost nobody will show you what the guardrail actually produces. Here is the record written by the run in the drawing at the top of this page — the real shape of it, with the names changed.
Three things this settles before they become arguments:
The record names the documents in scope. If an answer was wrong, you can tell immediately whether it was the reasoning or the retrieval — and those have different fixes.
The declined line matters more than the approved one. It is the proof that the fence exists and held.
A person, by name, at a time, with what they were shown. Accountability does not move to the software, and no regulator has ever accepted that it did.
An agent is a thin layer over whatever integration you already have. Where that layer is thin, the first job is not an agent at all — and we will say so before you spend anything.
This is the work KLM did for a decade before any of it was called agentic, and it is still most of what we are hired for.
One place the applications agree on: the model, the lineage, the access rules, and the answer to who owns which field. Most reconciliation pain is an architecture problem wearing a reporting costume.
Events, contracts and queues between the scheduling system, the ERP, the historian and the twenty-year-old application nobody is allowed to touch. Designed so a failure in one does not take the others with it.
A sequenced path with a working system at every step, costed before it starts. We have never recommended a rewrite where a strangler pattern would do.
A written answer to what to build, what to buy, what to retire, and where the operating cost lands in year three. Delivered as a document you can hand to a board, not a slide deck.
Every one of these ends with something written that survives the engagement: a decision, a running system, or an architecture your own people can maintain after we are gone.
We look at the actual work, not the org chart, and come back with a written answer on where an agent pays for itself here and where it will not.
One workflow, at level one or two, live — with its approval step, its evaluation suite and its ledger. Not a demonstration environment.
Standing architecture ownership for organisations that need the judgement without the headcount. The same person every month.
KLM is deliberately small. The architect you meet is the architect who writes the design, sits in the review, and is still reachable when something breaks eighteen months later.
Twenty-one years in software and enterprise architecture, built on work for American banks, insurers and energy companies. Landry founded KLM Consulting in 2015 and has led it since, taking the client work himself rather than brokering it out.
He currently serves as platform and enterprise architect across a portfolio of oil and gas engineering applications — owning the architecture end to end: the data platform beneath it, the integration between the applications, and the path from on-premises software to cloud services.
Before that he was principal technical architect on the dock and export platforms of a Houston midstream operator, built on .NET 8, Hot Chocolate GraphQL, Kafka, Elasticsearch and Azure Kubernetes Service, and used every day by the teams scheduling vessels and moving cargo. The worked example at the top of this page is drawn from that world.
Earlier came architecture and principal consulting roles at IBM / Neudesic, Improving Enterprises, Hitachi Consulting and CSC, delivering systems for banks, insurers and Fortune 500 operators across Texas and the Midwest.
Clients bring him the questions with no obvious answer: what to build, what to buy, what to retire, and where AI pays for itself rather than adding another subscription. He answers them in writing, with the operating cost attached.
MBA in management and finance, Avila University, Missouri. Doctoral research in business administration, in progress.
One paragraph is enough to tell whether this is an agent, an integration problem, or something you should not automate at all. You will get a straight answer either way, and it costs nothing to ask.