Question 1

Which model do you default to?

Accepted Answer

Claude Sonnet 4.6 or Opus 4.7 for anything involving tool use, longer reasoning, or production reliability. GPT-4 family when the customer is already locked into the OpenAI ecosystem. Open models (Llama 3, Mistral) when latency or per-call cost dominates the decision. We'll tell you the trade-off before we pick.

Question 2

How do you stop the agent from hallucinating?

Accepted Answer

We don't, at the model layer — nobody can. We constrain hallucination at the system layer: strict tool schemas, output validation, retrieval grounding with citations, human-in-the-loop for irreversible actions, and full audit logs so anything weird is traceable and reversible.

Question 3

Will it cost us a fortune in API calls?

Accepted Answer

Every agent we ship has a per-conversation cost dashboard. We architect prompts and retrieval to cap typical conversations at predictable cost (usually £0.01–£0.30 each, depending on the workflow). The pilot phase is specifically designed to measure cost before you roll it out broadly.

Question 4

Can the agent live inside our existing product?

Accepted Answer

Yes — as an in-app sidebar, a Slack bot, an email handler, a Telegram/WhatsApp/Discord bot, or a Twilio voice agent. Same agent core, different channels.

Question 5

Do we own the prompts and eval set?

Accepted Answer

Yes. Prompts, eval sets, tool definitions, and audit logs are all yours, in your repo, on your accounts. We don't keep a copy.

Question 6

What if the model gets deprecated next year?

Accepted Answer

Prompts are versioned, evals are reusable, and the model is a config value, not a hard-coded dependency. We've already migrated production agents from Claude 3.5 → 4 → 4.6 → 4.7 inside paid retainers without downtime.

Question 7

Can the agent take destructive actions on its own?

Accepted Answer

We default to confirm-first on anything that moves money, sends an external message, or writes to a production table. You can loosen that policy per tool, but it's a deliberate choice, not the default.

Question 8

What you won't build

Accepted Answer

Fully autonomous "AGI" agents that nobody supervises, agents that auto-execute irreversible legal or financial actions without a human in the loop, or chatbots whose only purpose is to look smart on a marketing page.

AI agents for e-commerce — every request is money.

What we ship for e-commerce teams

E-commerce-specific considerations

Other e-commerce-aligned services

Ready to scope a ai agent development build for e-commerce?