AI agents for business development work best on research, not relationships

Salesforce’s research found that sellers spend about 70% of their time on non-selling tasks, and I would guess business development sits no better. A large share of that time goes on reading filings, checking LinkedIn and building a reason to call, which is exactly the work AI agents for business development are good at.

My position is narrow and I will defend it. Give the agent the evidence and keep the person on the decision.

The usual mistake is to treat business development as one job and buy one tool for it, when the work splits into pieces that each deserve a different level of trust and a different owner.

Business development is four jobs with four different failure costs

Partner sourcing, account research, expansion triggers and relationship follow-up all sit under the same job title. Their failure costs are different, and that should decide how much autonomy each one gets.

A wrong account brief wastes ten minutes and a wrong expansion alert wastes a call, but a wrong partner approach can close a door with a company you will want to deal with in two years.

So I rank them by what a mistake costs rather than by what the tools claim, and the order that falls out is account research, then expansion triggers, then partner sourcing, then follow-up.

Account research is the safest task to hand off

An agent that reads a company’s site, recent news, job posts and public filings and returns a one-page brief is doing work no one enjoys and few people do consistently. I have written about the workflow in AI agent company research, and the BD version changes only the question the brief answers. For outbound the question is who to contact. For BD it is what this company is trying to do this year and who owns that goal.

Vendors such as Outreach publish agent recipes that follow the same pattern: gather signals, summarize, hand to a human. That is the shape I would copy.

What I would not do is let the brief drive the message. The agent finds the facts, and the person decides which fact is worth mentioning. A brief that says a company opened a Munich office tells you nothing about whether that office matters to the conversation.

Write the brief template yourself. Five fields are enough, covering what the company sells, what changed in the last 90 days, who owns the relevant budget, what they already use that touches your space, and what could not be verified. That last field matters most. An agent that is allowed to leave a field empty will leave it empty, and one that is not allowed to will invent something.

Partner sourcing needs overlap data for the agent and the fit call for you

Finding partners is a different problem, because the public web is a poor source for it. Account research works on public pages. Partner fit depends on whose customers overlap with yours, and that data sits in CRMs.

Crossbeam’s analysis of its own network found an average win-rate lift of 11.7% when partners were involved in a deal, and a much wider spread by ecosystem size. That is a vendor sampling its own users, so I read it as direction, not as a promise. It does tell me where the value is. Overlap between two customer lists is worth more than a long list of companies that merely look similar.

An agent can do two things well here. It can clean and normalize your account list so that matching works, and it can read a candidate partner’s integrations page, marketplace listing and case studies to say who they sell to, which gives you the raw material for a comparison with your own ICP.

It cannot tell you whether the founder across the table is someone you want your customers talking to. I would not pay for any tool that claims to score that.

If you have no CRM hygiene, start there. An agent running on an account list with duplicate companies and missing domains will produce a confident, wrong overlap report. I would sooner spend a week fixing the list than a month trusting the output.

Expansion triggers are cheap to detect and easy to misread

Watching existing accounts for hiring spikes, funding rounds, leadership changes and product launches is the most natural agent job in the whole set. It runs on a schedule, it reads public sources, and the output is a short list. I covered the logic in signal-based outbound, and the same logic applies to customers you already have.

The risk is the reading. A new VP of Sales can mean a bigger budget or a vendor cull. Funding can mean growth spend or a freeze on anything not tied to runway. The agent reports the event. Someone who knows the account has to say what it means.

My rule is that every expansion alert carries the event and its source link, plus the single hypothesis the agent thinks it points to. If the person who owns the account disagrees with the hypothesis more than half the time over a month, the agent is producing noise and the trigger list needs trimming.

Alerts without an owner go nowhere.

Follow-up is where I would keep the agent away from the send button

Relationship follow-up looks like the easiest win, since it is repetitive. It is also where an agent does its most visible damage.

An agent can log the call, pull out action items, flag that a partner went quiet six weeks ago and draft the next note. All of that is safe. Sending that note unseen is another matter. A partner who gets a generic check-in from someone they trust learns that the trust was thinner than they thought.

I would run drafts for a few weeks and count how often the owner sends them unedited. When that holds steady near all of them, widen the permission for low-stakes notes such as meeting recaps and calendar confirmations. Anything that proposes terms, mentions a competitor or follows a missed commitment stays manual permanently.

The strongest case for full autonomy, and why I still disagree

The best argument against my split is that humans are inconsistent, so a tireless agent that follows up on every partner will out-execute a busy one. I think that is partly true, and consistency is a real gain.

The evidence for autonomous agents in this kind of work is thin, though. Gartner expects more than 40% of agentic AI projects to be scrapped by the end of 2027, citing cost, unclear value and weak risk controls, and it warns about vendors relabeling chatbots as agents. Business development has the weakest feedback loop of any growth function, because a deal can take a year to close and a bad message may never be traced. An agent trained or tuned on that signal learns very slowly.

So the consistency gain is real but small compared with the cost of one badly judged message to a partner. Automate the preparation and keep the send.

If you are choosing what to automate across the whole business and not just BD, what to automate first sets out how I would rank the candidates, and best AI agents for business covers the tools by job. For BD specifically, my order is a research agent this month, a trigger watcher next quarter, and nothing that writes to a partner without a person reading it.

If you want a second opinion on where an agent belongs in your own BD motion, the free growth audit at app.momentumnexus.com is built for that, and the AI agents page shows what we run.