Key takeaways in 3 minutes
TX-1 Terminal Explorer is a prototype for agentic AI in high-accountability enterprise workflows.
The idea is simple: AI should not just explain operational failures. It should investigate them, propose a fix, validate the fix in a dry run, and then ask a human to approve the change before anything touches the system of record.
That approval gate is not a weakness. It is the product. In serious business environments, trust depends on evidence, accountability, auditability and controlled action.
The wider lesson is that enterprise AI needs to move from visibility to resolution without removing human ownership.
AI earns the right to act when it proves the work first.
The most useful AI in enterprise software may not be the AI that answers questions.
It may be the AI that notices something has broken, works out the likely fix, proves the fix in a safe environment, and then waits for a human to say yes.
That last bit matters.
In high-accountability work, autonomy is not automatically progress. Supply chain planners, finance teams, compliance analysts and operations leaders do not just need insight. They need controlled action, evidence, auditability and a clear answer to the awkward question nobody puts in the demo: who is responsible if this thing changes the wrong data?
That is the thinking behind TX-1 Terminal Explorer, a prototype I built to explore what agentic AI should look like when it is allowed near serious operational workflows.
The Problem Is Not Insight
Enterprise teams often know when something has gone wrong. The hard part is the recovery loop.
In supply chain modelling, a solver may return one blunt word: INFEASIBLE. Somewhere in the model, a demand forecast, capacity constraint, logistics lane or data assumption has made the scenario impossible.
What happens next is usually a manual hunt.
The modeller opens another screen. Then another. They check the source data, cross-reference rows, test a hypothesis, make a change, rerun the model, and hope the solver is in a better mood this time. If it fails again, the loop starts over.
That is expensive work for skilled people to be doing by hand.
The interesting AI question is not "can we summarise the failure?" It is "can the system close more of the recovery loop without stealing the decision from the human?"
From Visibility To Resolution
A lot of enterprise AI stops at visibility. It explains what happened, highlights an anomaly, or writes a tidy summary. Useful, yes. But visibility is not the same as resolution.
If the human still has to leave the product, find the records, work out the fix, test it manually, update the source system and document the decision afterwards, the AI has improved the commentary rather than the workflow.
TX-1 explores a different pattern.
The system classifies the failure, inspects the relevant data, proposes a patch, runs that patch inside a rolled-back transaction, reruns the solver against the temporary change, and only then presents the human with an action card.
The user is not being asked to approve a guess.
They are being asked to approve a validated change.
The Approval Gate Is The Product
Human-in-the-loop AI is often framed as a compromise. The AI is not trusted enough to act, so the poor human must sit there like a nervous adult at a school disco.
I think that is the wrong framing.
In high-stakes environments, the approval gate is not a patch over weak AI. It is part of the product value.
The user needs to know what will change, why it will change, what evidence supports it, what happens if they do nothing, and whether their decision will be recorded. The organisation needs to know that an accountable person approved the state change.
This is not just a UX detail. It is architecture.
In TX-1, the agent can propose. It can validate. It can prepare the action. But it cannot commit the change until the human approves. The graph pauses. The database is untouched. The audit trail waits for a decision.
That makes uncontrolled action structurally impossible rather than politely discouraged in a governance document nobody reads.
Dry Runs Build Trust
Trust in AI does not come from a confident tone. It comes from evidence.
Before TX-1 shows a proposed fix, the system has to prove that the fix works in a dry run. It constructs the change, applies it inside a transaction, reruns the solver, checks that feasibility is restored, and rolls the change back.
Only then does the action card appear.
That changes the user experience completely. The human is not forced to mentally simulate whether the AI has got it right. The system has already done the verification work. The human can focus on judgement: should this proven fix be applied in this business context?
That is a better division of labour.
The machine does the repetitive investigation, validation and comparison. The human owns the decision.
The Interface Has To Tell The Truth
I chose a terminal-style interface for TX-1 because the product is about action, not passive reporting.
A dashboard would have been the obvious enterprise choice. Cards, charts, filters, a neat little panel that says "AI Insight" as if the insight has arrived wearing a lanyard.
But dashboards are mostly built for observation. TX-1 needed to feel closer to an operating surface: a place where intention becomes execution, where every agent step is visible, and where the user can see the reasoning stream before a decision is requested.
That is why the interface is sparse, dense and direct. It does not try to make serious data changes feel fluffy. It shows the agent sequence, the proposed patch, the before/after impact and the approval state.
The aesthetic is not decoration. It is part of the trust model.
What TX-1 Demonstrates
TX-1 is a prototype, not a production product. The supply chain scenario is synthetic. The company is fictional. The model is deliberately tractable.
But the design pattern is real.
It applies wherever a system detects a problem, reasons about a response, validates that response, and needs an accountable person to approve the change before it touches the system of record.
That could be supply chain constraints, financial thresholds, compliance exceptions, data quality failures, support routing, operational planning or any workflow where the cost of unchecked automation is higher than the cost of a short approval.
The supply chain scenario is the vehicle. The larger argument is about enterprise AI trust.
AI earns the right to act when it can show its work, prove its recommendation, wait for approval, and leave a record behind.
The Practical Move
Take one AI opportunity in a serious workflow and map the recovery loop.
Ask:
- What failure or signal starts the process?
- What investigation does a skilled human do today?
- What data does the system need to inspect?
- What change could the AI propose?
- How could that change be validated before it is shown?
- What should the human approve, reject or adjust?
- Where does the approved change need to land?
- What audit trail would make the organisation comfortable?
If you cannot answer those questions, you may have an AI feature. If you can, you may have the beginning of an AI product architecture.
The useful future is not AI that acts like a confident intern with database access. It is AI that does the hard preparation, earns the decision, and then waits for the person who is accountable.
- Sketchy black line editorial image on warm off-white background with one accent colour: {{BLOG_ACCENT: Amethyst Smoke #947EB0}}. A supply chain planner reviews an AI action card with a large approval button while the AI shows a dry-run receipt.
- Hand-drawn workflow loop: solver failure, agent diagnosis, dry-run validation, human approval, committed fix, audit trail. Use {{BLOG_ACCENT: Amethyst Smoke #947EB0}} only for the approval and audit points.
- Two-column comparison: "AI insight" shows a summary dumped on a desk; "AI resolution" shows validated fix, evidence, approval and system-of-record update.
- Metaphor image: an AI politely holding a spanner and a clipboard, waiting outside a locked database door while a human checks the evidence. Dry humour, sparse office-cartoon energy.
"Visibility is not the same as resolution."
"The user is not being asked to approve a guess. They are being asked to approve a validated change."
"The approval gate is not a patch over weak AI. It is part of the product value."
"Trust in AI does not come from a confident tone. It comes from evidence."
"AI earns the right to act when it can show its work, prove its recommendation, wait for approval, and leave a record behind."