Insights

What does it take to scale AI? Not set and forget

Nov 13, 2026, 6:30:00 AM · Rick Bawcum

← All insights

The assumption that breaks at scale

Most software is deterministic. Test it once, and it does the same thing tomorrow until someone changes the code. Organizations built their habits around that: validate before launch, sign off, move on.

AI tools do not work that way. They are non-deterministic. The same question can get different answers on different days, and the answers can move even when nobody on your team touched anything. A vendor updates the model. A staff member edits the instructions. A new document lands in the knowledge base. A one-time validation describes one day. It does not describe the day a member asks.

That matters most when you try to scale. Scaling AI means more members relying on it, more questions, more vendors and more places it is embedded. Each of those multiplies the ways it can drift from what you approved. Scaling the use without scaling the checking is how a small problem becomes a member-facing one.

What can change underneath you

Here are the threats that a launch-day test cannot cover. The categories follow public frameworks such as the OWASP Top 10 for LLM applications and the NIST AI Risk Management Framework.

  1. Model drift. The vendor updates or swaps the underlying model, or routing changes, and behavior shifts without notice.
  2. Configuration drift. Instructions, settings, tone rules and escalation rules get edited over time, one reasonable change at a time.
  3. Content drift. The knowledge base goes stale. The assistant keeps answering confidently from last year's deadline, fee or policy.
  4. Prompt injection, direct. A user types instructions meant to override the assistant's rules.
  5. Prompt injection, indirect. Instructions hidden in a web page, document or email that the AI reads can redirect what it does.
  6. Data leakage. The assistant reveals member or member-company information, or its own instructions, to someone who should not see them.
  7. Jailbreaks. Phrasing that talks the assistant out of guardrails it held last month.
  8. Hallucination. Plausible, wrong answers about events, dates, eligibility or credentials.
  9. Excessive agency. An agent that can take actions, such as sending email or changing records, takes one it should not.
  10. Poisoned or manipulated sources. Content the assistant retrieves from has been altered, deliberately or by accident.
  11. Inconsistency across channels. The website chat, the help desk and the event platform each give a different answer to the same question.
  12. Vendor and dependency change. A vendor changes terms, features or subprocessors, or retires a model version you depended on.

Not every item applies to every assistant. A simple FAQ bot has a short list, and an agent with system access has a long one. The point is that most of these appear after launch, not before it.

What scaling actually takes

Treat it as an operating practice, not a launch milestone:

  • Write down the guardrails, including what the AI must never say, which data it must not touch, and what goes to a person.
  • Keep a fixed set of real questions and run it from the outside, the way a member would ask.
  • Re-test regularly, and again after any known change: a vendor update, a staff edit, a complaint.
  • Include adversarial questions, such as injection and leakage attempts, not only friendly ones.
  • Keep dated results, so a change shows up as a difference rather than a feeling.
  • Name an owner who acts when a result comes back bad.

A small team can run a modest version of this itself, and that beats nothing. What tends to lapse is the regular, independent part.

Where Cimbiotic fits

CimAssure is built for this. Versioned tests run against your chatbots and agents from the outside, on a regular schedule and when the AI changes. The results come back as a dated Assurance Statement your Board can check: evidence of how the AI behaved when we tested it, and one input to your own oversight. It is not a certification, audit or attestation.

Choose your next step:

Sources

Where does your organization actually stand?

Seventeen questions, about five minutes, no sign-up to see your result.

Take the free Scorecard Book a 30-minute call