Stopping AI hallucinations in tender responses

There's a thing happening in UK procurement that nobody wants to admit out loud.

Evaluators are reading AI-drafted bids. They can usually tell. And the ones with fabricated content — invented case studies, made-up certifications, contracts that never happened — are increasingly getting noticed.

A central government procurement lead told me last month: "We had a bid this quarter that listed a £4m project with a buyer who'd never heard of the company. They didn't even know how badly they'd burned the relationship."

The problem isn't AI. The problem is AI without a grounded knowledge base. Here's the difference, and how to use AI safely.

How AI hallucinations happen in bids

Large language models predict plausible-sounding text. When you ask one to "write our experience delivering NHS contracts", and the model doesn't know your experience, it generates something plausible: a project name, a buyer, a value, an outcome. None of it is real. All of it sounds professional.

Three particular hot zones:

1. Past performance / case studies — most common hallucination
2. Specific accreditations — model invents ISO numbers, frameworks, certificate bodies
3. Named team members — invents people, sometimes with realistic LinkedIn-style backgrounds

The evaluator now has a bid where everything reads well and some of it is fiction. When they spot one fabrication (and they will, because they verify), they treat the whole bid as potentially misrepresented. You're out — and on a list for next time.

The fix: grounding, not generation

The principle is simple: the AI doesn't generate from nothing. It generates from your real documents.

In practice that means:

1. Build a knowledge base of your actual capability: past contracts (with real buyer names, values, dates, outcomes), accreditations (with real numbers and expiry dates), team CVs, case studies.
2. The AI gets your knowledge base as input every time it drafts.
3. The AI flags assumptions it had to make — gaps it papered over with reasonable-sounding text — so you can fix them before submission.

When done right, an AI-drafted response cannot fabricate a case study because the prompt explicitly says "only use the case studies provided." It will leave the answer half-finished if the knowledge base doesn't have what's needed — which is the right behaviour. Better a flag than a fiction.

Practical guardrails when AI-drafting your next bid

If you're not using a grounded tool, here are the manual guardrails:

1. Strip every specific claim from the draft and verify

Read the AI output. For every concrete claim — a contract name, a number, a date, a person, a certification — check it against your records. If you can't verify it, delete it. Don't tweak it.

2. Never let AI generate past performance

This is the hallucination vector that costs bids most. Past performance must come from your own records, not from generation. Use AI to polish a case study you've written yourself; don't let it create one.

3. Treat accreditation numbers as sacred

The model will happily invent a Cyber Essentials Plus certificate number. Buyers check. If the number on your bid doesn't match the IASME register, you're done.

4. Get a human-readable diff

Before submission, do a side-by-side read of the AI draft against your knowledge base. Every claim in the draft should map to a source. Anything that doesn't — delete or rewrite.

5. Use AI for the structure, not the content

The strongest use of AI in bid writing is shaping the answer:

• Suggesting which sections to lead with
• Restructuring a wall of text into the buyer's preferred Situation-Task-Action-Result frame
• Catching missing compliance signals against the published scoring matrix
• Tightening verbose paragraphs

These are all safe because they operate on text you've already written and verified.

What evaluators are now watching for

The red flags evaluators have learned to spot:

• Case studies with buyers the evaluator's network has never heard of
• Round numbers that look suspiciously generated (£500,000, exactly 100 staff, "23% improvement")
• Generic outcomes that could fit any contract ("delivered improved efficiency")
• Accreditation numbers that don't match public registers
• Inconsistent tone between sections (one written by the model, one by a human)

If your bid has any of these patterns, expect a verification call. If the verification fails, expect to be excluded.

The 80/20

Use AI. Don't fight it. But:

1. Ground every AI draft in a knowledge base of your real capability
2. Verify every specific claim before submission — buyers, numbers, dates, certifications
3. Use AI for structure, not for past performance generation

That's the line between "AI that wins you more bids" and "AI that gets you excluded from frameworks".

This is what TenderForge does by design — every draft is grounded in your knowledge base of real contracts, accreditations, team CVs and case studies. Where it has to assume something, it tells you in the draft so you can fix or delete it before submission.