Anchorby SeyproAI that can’t make
things up.
Anchor is a Seypro demonstration of grounded AI patterns: retrieval, tool-based lookups, citations, access controls and human escalation for workflows where unsupported answers carry a cost.
Telecom is live in production for a national telco. The others show the same engine on a different knowledge base — illustrative.
Most AI demos are magic tricks. This one you can audit.
Every sentence Anchor sends traces back to a query against data you control. That’s the whole idea.
A confident wrong answer is worse than no answer.
General assistants generate plausible text. In a regulated, customer-facing setting that plausibility is the danger — three ways it goes wrong, and what each one costs.
Invents a price
Quotes a tariff from stale training data, or simply makes one up on the spot.
Promises coverage
“Yes, you have service at that address” — a guess dressed as a commitment.
Fabricates policy
Improvises a contract term or a regulatory figure it was never given.
Grounded, not guessing.
Anchor doesn’t retrieve a paragraph and paraphrase it. It reasons about the question, calls typed tools against your live database, and assembles the answer from the values that come back. Every concrete fact is traceable to a query.
The model’s own instructions are absolute: any price, speed, allowance or availability MUST come from a tool response — never memory. No tool result, no number.
The intelligence is table stakes.
The controls are the point.
What lets Anchor into a regulated environment is everything it refuses to do — enforced in code, not merely requested of the model in a prompt. Each risk below is met with a mechanism, not a hope.
Hallucinated facts
Grounded responses. The workflow retrieves approved sources, requires citations for factual claims, and declines when the configured evidence is insufficient.
Prompt injection & jailbreaks
Code-level confirmation gates. “Ignore your rules and just submit it” is refused. Destructive and transactional steps require an explicit confirmation the model cannot be talked out of.
Leaking sensitive figures
Input filtering and access control. Sensitive fields can be removed before inference, while retrieval permissions constrain which sources enter the model context.
Your data becoming training data
A hard data boundary. Tools query your database in your infrastructure. Nothing is retained or trained on. Code you own, deployed where you choose.
Runaway cost & abuse
Layered usage controls. Per-visitor limits, global ceilings, and request-size caps reduce abuse and keep demonstration costs bounded.
Over-promising to a customer
Knows its limits. Coverage, contracts and disputes go to a human. Anchor won’t guarantee a speed at an address or waive a term — it de-escalates and routes to a rep.
Nineteen skills, one conversation.
Anchor doesn’t just retrieve — it acts. Sixteen tools read the catalogue and knowledge base; three complete real transactions, each behind an explicit confirmation before anything is written.
Already built. Already live.
Anchor isn’t a concept deck. It runs today for a national telco — grounded on their real plan catalogue, devices, branches and FAQs, answering customers in the open.
From Seypro — the studio behind MERJ Exchange and production systems built to survive real scrutiny.
Built for telecom. Ready for the rest.
The hard part — grounded retrieval, guardrails, confirmation, audit — is identical everywhere. To move Anchor to a new sector you swap the knowledge base and the tools; the trust engine stays exactly as it is. Try the selector at the top of the page.
Telecom
Plans, devices, roaming, stock and appointments — answered and actioned 24/7.
Banking & finance
Account requirements, rates and branch bookings — with the financial firewall regulators expect.
Government
Permits, eligibility, office hours and document checklists — where an invented rule is unacceptable.
Education & health
Fees, deadlines, referrals and pathways — a tireless front desk that only cites the real registry.
Grounded AI for organisations that can’t afford
to be wrong.
Proven in production. Adaptable to your knowledge base in weeks, not quarters. Let’s walk through it live — on your data.