Data agents that watch the data, flag what moved, and explain what is driving it
An agent is only as good as the layer beneath it, so we build the governed data foundation first and put the model on top. The agent then monitors metrics, explains what moved them, and with approval triggers the operation.
- Engagement shape
- Week 1 audit, weeks 2-6 build
- Team on it
- A pod of three
- Median to first production-grade artefact
- Six weeks
- Partner status
- Anthropic Build Partner
- Track record since 2023
- 50+ projects, 20+ clients
What happens between the data moving and someone acting
Five steps, and most tools stop after the second.
- Watch
The agent reads the data continuously
Metrics are monitored as they move, rather than waiting for someone to open a dashboard on Monday morning. The agent holds the current state, not last Friday's extract.
- Flag
A signal falls outside the norm
The deviation is raised in real time instead of surfacing long after the fact in a review meeting, when the decision it should have changed has already been taken.
- Explain
It shows what is driving the movement
Not only that a number moved, but which segments, products or accounts moved it, and whether a chart helps here or a single sentence is enough.
- Recommend
It proposes the decision
Named options with reasons attached: shift a work order to the night shift, order the missing steel and plastic, cut machine utilisation to 85 percent.
- Act
It triggers the operation
With approval, the agent raises the ERP purchase request, sends the alert or starts the campaign, and what it did is recorded alongside the signal that prompted it.
Six things the agents we build actually do
Each one is a capability we implement, not a mode you switch on.
Track performance metrics in real time
Continuous monitoring rather than a scheduled assembly job. The agent knows the state of the business now, and can say when it last checked.
The foundation this needs →Warn when a signal leaves the norm
The agent flags a deviation as it happens. In our own build it named an anomaly in one segment and raised it before anyone had opened a report.
Reveal what is driving the trend
The follow-up question is the useful one. The agent breaks a movement down to the segments and the underlying records behind it.
Guide the decision
It answers a question like whether there is enough steel and plastic for production, and if not, prepares the orders rather than reporting the gap.
Launch operations or campaigns
Purchase requests, alerts and workflows are triggered from the data the agent analysed, with the approval step kept where your process wants it.
Answer in plain language
The question is taken as asked, in a sentence, and answered without SQL, jargon or another tool for people to learn.
The governed layer we build on →What makes an agent proactive rather than another chat box
Four tests, and an agent that fails them gets quietly abandoned.
It comes to you with the answer
The agent starts the conversation when the data warrants it. Waiting to be asked means it only ever finds what someone already suspected.
It knows when a chart helps and when a sentence is enough
Most answers are one number and one reason. Rendering a visual for that is theatre, and it slows the person down.
It understands the question and replies without jargon
If the answer needs translating before it can be forwarded, the agent has moved the work rather than removed it.
It can act, not only report
Analysing, explaining and then taking the action is the difference between checking a report and having the decision in front of you.
Why we do not just point a model at your database
The failures on the left are properties of the data layer, not of the model.
- The model reads whatever the source systems happened to sendA cleaned and joined layer the agent reads instead
- Two teams hold two definitions of the same metricOne agreed definition the agent is bound to
- Security and governance are handled tool by toolGovernance applied once, in the layer the agent queries
- A confident wrong answer looks exactly like a right oneA governed layer built to prevent hallucinations and hold answer quality
Figures from our reference implementation
These come from our own build, not from a named client account, and you should read them that way.
The path a question takes
One question, timed end to end on existing infrastructure.
- 0s
The question is asked
In plain language, in whichever surface the person already works in. No new tool, no chart planned in advance.
- 0.4s
Data fetched
The agent reads the infrastructure that already exists. No new pipeline is built to answer this particular question.
- 0.8s
Processing
The result is cut against the definitions the business agreed, so the answer matches what Finance would have produced by hand.
- 1.2s
Answer returned
Against the two to three week engineering sprint the same open-ended question would otherwise have queued behind.
The same path, drawn to scale
Bar length is how far into the second each stage of one question lands.
How an agent engagement runs
The same three phases we use across every engagement.
- Week 1
Fixed-fee audit
Two calls and a one-pager you keep either way: what the agent would need to read, whether the layer underneath can support it, and whether an agent is the right answer at all.
- Weeks 2-6
Build
A pod of three. Across our engagements the median to a first production-grade artefact is six weeks, and an agent is held to the same bar.
- Week 7 onward
Operate
The agent runs against live data, its answers are monitored, and the definitions it reads are kept in step with the business as they change.
Where the audit sits in the six weeks
One marked moment in the window between the first call and a first production-grade artefact.
Where a data agent is the wrong answer
Four cases where we would rather say so in week one than once the build is under way.
The layer underneath is not ready
If the systems disagree with each other and nothing has been reconciled, an agent will answer confidently and be wrong, which is worse than no answer. The honest first project is data engineering, and the agent comes after it.
Nobody is asking the questions in the first place
If your existing reports go unopened because the organisation is not making decisions from data, an agent gives people a faster way to not ask. That is a management problem and we cannot sell you a fix for it.
Every action already requires a signature
The agent can prepare the purchase request or the schedule change, but if your process needs a named human approval on each one anyway, most of the gain is in the analysis rather than the automation. Worth pricing it that way.
You want the chat box, not the foundation
We do not plug a language model into a warehouse and call it an agent. If the budget covers the interface but not the governed data layer beneath it, we will decline the work rather than build something that quietly fails.
The dashboards nobody opens still cost someone a morning
Recurring reports accumulate faster than anyone retires them.
Questions buyers put to us about agents
The five that come up in almost every first call.
Can you put an agent on the warehouse we already have?
Often yes, and that is the cheaper path when the estate is sound. The agent reads existing infrastructure, so no new pipeline is required to answer a question. What we check in week one is whether the definitions are agreed and whether governance can be enforced on the layer the agent queries. If those two hold, we build on what you have. If they do not, fixing them is the first piece of work and we will say so.
How do we know the agent is not making things up?
Two things, and neither is the model. First, the data layer: the agent reads a cleaned, governed set of definitions rather than raw tables, so there is one answer to what a metric means. That is what keeps hallucinations out. Second, monitoring: the agent's own accuracy, latency and drift are watched in production the way any other model would be, so quality is visible rather than assumed.
Which tools does it plug into?
The agents plug into the tools and workflows already in use rather than adding a place to visit. We have built agents in Copilot Studio, including multi-agent setups where one agent triggers another, and agents that raise purchase requests directly in an ERP. The integration surface is decided in the Week 1 audit against what you actually run.
Does this replace our dashboards?
Some of them, and you should want it to. In our reference build fourteen recurring reports were retired because the questions behind them were better answered on request. Dashboards that carry a decision, that people genuinely open, are worth keeping. The rest are the analytical debt an agent is good at clearing.
What does it cost?
We do not publish a price for agent work, because the cost is driven almost entirely by the state of the data underneath rather than by the agent itself. The Week 1 audit is fixed fee, and it ends with a one-pager telling you what the build would involve. If the answer is that you need data engineering before you need an agent, that comes out in week one at audit cost rather than once a build is under way.
The rest of the practice
These are genuinely different jobs with different ways of failing. Most engagements start in one of them.
- Data engineeringJobs that finish, migrations that reconcile, and storage that stops paying for cold data.
- Data integration and governanceOne set of records the finance team and the operations team both accept.
- Data platforms and modernisationOff the platform you outgrew, without a twelve-month freeze on new reporting.
- Apache SupersetSuperset built and run by people who commit to the project, including embedding and Kubernetes.
- Command centresThe one screen an operations floor runs the day from, not another dashboard.
- AI audit and roadmapOne week, and you know which two or three AI initiatives are worth building.
- AI evaluationA graded eval set, because an AI system cannot simply pass or fail a test suite.
- AI governanceGovernance you can defend in a board meeting and audit on demand.
- AI agentsOne reasoning agent on Claude, on your live data, with evidence and a kill switch.
- AI use case, guaranteedOne production agent in eight weeks, or you do not pay for the build.
- Applications and automationThe system your team works in all day, built or replaced in slices.
- Application modernisationThe system nobody wants to touch, replaced a slice at a time rather than rewritten.
- System integrationSystems that stop disagreeing about the same customer, order and item.
- Process digitisationThe process that still runs on paper, WhatsApp and one shared spreadsheet.
Start with the audit and find out whether an agent is the right answer
Two calls, a fixed fee, and a one-pager you keep either way: what an agent would need to read, whether your current layer can support it, and what the first version would do. If the foundation comes first, we will say so.