Mindset Stories Our Focus Security & Governance Get in touch
All stories
Energy & Gas · Poland

Unimot Energia i Gaz frees its specialists
for the bids worth winning

An AI tender automation workflow runs the compliance review on every public tender,
drafts the formal questions, and hands a ready package to the specialist for sign-off.

<$1
model cost per tender
100+
point compliance checklist
<20 min
email arrival to analysed
<3 mo
kickoff to production

The challenge

Expert time, spent on the routine.

The offering team spent its week on routine, replicable work - the same compliance review on every tender, the same questions to draft, the same document structures to verify. They knew exactly what to look for; the work itself was mechanical.

The brief came from C-level: place AI at the heart of the offering workflow, free human potential for higher-value work, keep them owning every decision that leaves the company.

The solution

Seven stages. One human gate.

We built an AI workflow that ingests incoming tenders, runs a 100+ point compliance review on each one, auto-drafts the formal questions, and routes everything to a specialist for approval. It runs in seven stages, from email arrival to a ready-for-review package; each stage is owned by a specialised AI agent, and the whole flow sits behind one human-controlled gate.

1

Tenders arrive

Incoming tenders land by email - often several in a single message.

2

Profile filter

Profile-mismatched tenders are filtered out before any model cost is spent on full analysis.

3

Read & structure

Each tender is read and broken into the structures the review needs.

4

100+ point compliance review

Every tender is checked against a 100+ point compliance checklist.

5

Confidence-banded verdicts

Every verdict carries a confidence band, so the specialist sees how sure the model is on each check.

6

Auto-draft the questions

Where the document doesn't carry an answer, the system says “I don't know” and drafts a formal question to the contracting authority.

Specialist sign-off

Nothing leaves the system until a specialist approves it - the gate is built into the system, not just a rule.

Governance & compliance

Designed to ask when it isn't sure.

Runs on the client's own infrastructure

The workflow runs on Unimot Energia i Gaz's own infrastructure, under their direct Anthropic and OpenAI subscriptions - no third-party SaaS sits between the company and the models.

Public-domain inputs only

Every document the workflow reads is a published tender from public portals. No PII enters the workflow.

One structural human gate

No formal question and no compliance verdict leaves the system without specialist approval - that gate is built in, not a policy that can be waived.

Per-call audit log

Every model invocation is recorded - provider, model, time, cost - and is replayable months later for any audit or regulator request.

Results

A typical batch, processed while no specialist watched.

Take a typical batch - several tenders arriving in a single email, profile-mismatched ones filtered out before any model cost is spent on full analysis.
The system processed the rest on its own:

  • Under 20 minutes end-to-end, from email arrival to all tenders analysed and questions drafted.
  • Hundreds of compliance checks - each gap turned into a formal question for the specialist to send.
  • Under $1 per tender in model cost.
  • Zero specialist intervention during processing - specialist time is spent only at the approval gate.

What the numbers don't show

  • The offering team moves up the value chain. Repetitive checklist work is gone - specialist time now goes into the work that actually needs an expert, reviewed against confidence-banded verdicts instead of an opaque model output.
  • Every run is auditable. A regulator, an internal auditor, or a future bid review can reconstruct every decision the system made, what the model saw, and what the specialist approved.
  • The results stay with the client. The workflow runs on Unimot Energia i Gaz's own infrastructure; the data and the system both live in the client's environment.

Volume to the AI. Judgment to the specialist. That's the split that scales any regulated,
document-heavy workflow. That's governed AI, already in production.

Questions

Can a public tender workflow be automated end-to-end with AI?
Not without a specialist in the loop, and not honestly. The system automates the reading, the compliance review, and the question-drafting; every output routes to a specialist for approval. The specialist still owns the final decision.
How do you keep an LLM from hallucinating its way through a tender?
Three controls. The model returns “I don't know” instead of guessing when the document doesn't carry an answer. Every verdict carries a confidence band. And every formal question stays in draft until a specialist approves it.
Does the system replace the offering team?
No. It removes the repetitive compliance work and lets the team spend specialist time on the work that actually needs an expert. AI sits at the heart of the process; the specialist sits at the top of it.
Where does the data live, and what leaves the company?
The workflow runs on the client's own infrastructure. Inputs are public-domain tender documents - no PII. The only outbound traffic is the model calls themselves, sent directly under the client's own Anthropic and OpenAI subscriptions.
What does an AI tender pipeline cost to run per tender?
Under $1 in model cost on a typical run. Heavier tenders cost more. Profile-mismatched tenders are filtered out before full analysis, and prompt caching absorbs the bulk of inference cost on long documents.
How long does a deployment like this take?
Less than three months, end-to-end from kickoff to production-ready.

Got a queue of repetitive, document-heavy reviews?

Book a discovery call