← Insights

Internal copilots that actually get used

Most internal AI assistants are abandoned within a month. The ones that survive have a narrow job and live where the work already happens.

Sarthak Jain, Co-founder · · 5 min read

The pattern repeats. A company builds an internal assistant, launches it with a chat box and a "ask me anything about the business" prompt, sees a spike of curious usage, and watches it fall to near zero in about four weeks.

The failure is rarely the model. It is the framing.

"Ask me anything" is not a product

A blank prompt puts the entire burden of figuring out what the tool is good for onto the person using it. They try three things. One works, two do not. They stop trusting it, and they stop coming back.

The assistants that survive do one narrow, nameable job. "Draft the first-pass reply to this support ticket." "Find every clause in this contract that mentions termination." "Summarise what changed in this account since last quarter."

Narrow scope is what makes the output judgeable. A user can tell immediately whether a draft reply is usable. They cannot tell whether a general answer about the business is right.

It has to live where the work is

The second failure is location. If the assistant lives on its own internal site, using it requires a deliberate decision to leave the tool you are currently in. That decision loses to habit almost every time.

Put it inside the CRM, the inbox, the ticketing system, the editor - wherever the work already happens. Adoption is mostly a question of proximity.

Retrieval quality beats model choice

When an internal copilot gives a wrong answer, the cause is usually not the model. It is that the right document was never retrieved, or the source of truth is three conflicting documents and nobody has said which one wins.

Which means the useful work is often unglamorous: deciding what the canonical source is, removing the stale copies, and making sure the retrievable set matches what a person would have consulted.

An assistant built over a messy knowledge base does not fix the mess. It makes it queryable, and confidently wrong.

Measure refusal, not just usage

Usage counts flatter you. The two numbers worth watching are how often people accept the output unedited, and how often the assistant declines rather than guessing.

A copilot that says "I do not have a source for that" is trusted. One that always answers is used once.

If this describes your situation, start with a baseline.

We document where you stand before recommending anything - so whatever we claim at 90 days has a before to be measured against.