What is AI actually doing in your business?

AI is already doing real work in most firms. It drafts proposals, updates CRM records, writes code and answers client emails. We want to know where that’s happening and what it actually produces.

Our guess is that this list gets longer every month in a lot of firms. BDC’s new study on SME digital transformation has a message about which firms get the most out of it. (source)

Connect systems and train people

Among SMEs with fully integrated data, 94% are satisfied with the return on their AI investment. With partly integrated or unintegrated data, 63% are satisfied with that return. (source) SMEs whose AI draws on their own internal data are also more satisfied (85%) than those using only external data (64%). (source)

86% of SMEs that train their staff in AI are satisfied with the return, against 53% of those using AI without any training. (source) BDC calls training “a key to maximizing the return on AI investment.” (source)

AI does more for a firm when it is connected to that firm’s own systems and its people know how to use it.

Almost half of SMEs with digital data (47%) say it isn’t integrated at all or only very partially. (source) About half (51%) gave their teams AI training in 2025. (source) There is a lot of room to grow here.

After integration and training

Once a firm has done that work, AI is in the CRM, the shared drive, the codebase, and the inbox. Many people use those tools for small pieces of work every day.

Putting AI in those tools is the point of integrating. It is also when it gets hard to see what AI is actually doing. Each tool has its own view, and none of them shows the work as a whole.

Judge the finished work

We look at the finished work. That is what the firm is accountable for, and it is where quality shows up. The proposal that went out and the commit that shipped are part of that work, and so is the email the client read.

The agent’s trace of how something got made is useful for debugging. The trace includes the prompts and the tool calls, along with the retries. It doesn’t tell us much about whether the result was any good.

BDC’s guide on creating value with AI suggests tracking benefits like “average time saved per task,” along with “the number of errors eliminated and improvements in customer satisfaction.” (source) That’s a good list. For a services firm we would tie it to the deliverable. Time saved on a proposal means a lot more when the firm also knows the proposal held up.

BDC doesn’t break its results out by industry, so in this part we are applying the general SME findings to the firms we spend most of our time with.

A worked example

Take a made-up 25 person consulting firm. It has connected AI to its CRM and shared drive, and trained the team. AI now drafts first versions of proposals from past work, and updates the CRM after client calls.

A couple of months in, the partners look at the finished proposals instead of the usage reports. The firm background and case study sections go out with light edits. The scoping and pricing sections get rewritten almost every time. In the CRM, the call notes hold up, but the next-step fields often need fixing.

A usage dashboard would not show them this. Drafting from past work is holding up, and scoping is not holding up. Their hunch is that the AI can’t see what past jobs actually took. They test that hunch by connecting the project history for the next few proposals and checking whether those finished scopes need less rewriting. If they do, that is the next connection. If they do not, the firm keeps looking. They add no new metrics, and they look at the work.

Where to connect next

Firms integrate and train so AI does real work. They see where it shows up and what it produces, and they use that to pick what to connect next.

The angle we’re taking at Sabia is AI spend vs Work Produced. We start from the finished work and see where AI showed up in it. We judge the quality on that work, and we do not use the agent’s trace for that judgment.