An AI sales agent can only recommend what its inputs let it see, and almost none of them can see who in your company already knows the buyer. They read CRM fields, email activity and intent data. They do not read the relationship graph: the executive who worked alongside the target CFO for years, the customer champion who sits on the same board, the investor who backs the buyer's last company. So when an agent is asked how to reach someone, the only answer available to it is another cold touch.
Force Management puts the problem simply in its AI ebook: AI tools are "only as good as their inputs". This entry looks at the input much of sales AI is missing, what it looks like when it is supplied, and how to plug it into the agents you already run.
Why the missing input matters now
Gartner predicts that by 2028 AI agents will outnumber sellers 10x, yet fewer than 40% of sellers will report that AI agents improved their productivity (Gartner, 18 November 2025). That gap is not a model problem. The models write well. It is an input problem. An agent that sees the same fields as every other vendor's agent produces the same recommendations: find more contacts, send more sequences, follow up faster. Multiply that by ten agents per seller and buyers are likely to see more near-identical cold outreach, which gives them little reason to reply. See AI agents will outnumber sellers 10x for the full prediction.
The one thing that reliably cuts through is a person the buyer already trusts. That fact sits outside every system these agents read.
What common AI sales agents see, and what they miss
| Agent type | Inputs it sees | What it misses |
|---|---|---|
| AI SDR / outbound agent | Contact databases, firmographics, intent signals, sequence performance | Anyone in your company who could introduce the rep instead of emailing cold |
| Research and account-brief agent | Public web, news, filings, CRM account history | Which of your executives, investors or customers knows the people in the brief |
| Conversation intelligence agent | Call recordings, transcripts, talk ratios, topics raised | Relationships with stakeholders who never joined a call |
| Forecasting and deal-inspection agent | Stage, amount, close date, activity counts, email engagement | Whether the deal has a warm path to the economic buyer, or only one thread |
| Email and follow-up assistant | The rep's inbox and calendar | Relationships held by everyone else in the company |
| Coaching agent | Product content, messaging, call reviews | Who the rep should be talking to, as opposed to what to say |
Every row has the same blind spot. Each agent is bounded by one person's inbox or one system's fields. None of them can see the combined network of your executives, employees, board, investors, advisors, customer champions and partners, which for many companies is a go-to-market asset nobody has mapped.
Why agents default to cold volume
An agent optimises for the outcome it can influence with the levers it has. If its only levers are contacts and messages, the path to more meetings is more contacts and more messages. That is not bad design; it is a rational response to incomplete inputs.
Give the same agent relationship data and the options change. Instead of "add these contacts to a sequence", it can say "your VP of Engineering spent years working with this CISO; ask her for an introduction". The recommendation shrinks from a list of cold contacts to one named person, one relationship and one ask.
What relationship data looks like as an input
Relationship data is not a list of LinkedIn connections. To be useful to an agent it has to answer three questions: who knows this buyer, how well, and who inside the company should make the ask.
- Signals. Boomerang monitors more than 80 relationship signals continuously: work overlap, shared universities, past meetings, email and calendar history, shared patents, co-authored papers, public testimonials, social engagement with target accounts, shared event stages, board seats, portfolio overlap and mutual connections. See relationship signals.
- Four connector types. Executives and employees; investors, advisors and board members; customer champions; and partners. Each has a different cadence for how often they can be asked.
- Path strength. A path score of capability times willingness, per connector, per deal. It separates a strong relationship from one someone will actually act on. See connector score.
- The internal hop. The person inside your company with the strongest tie to that connector, so the ask goes out from whoever is most likely to get a yes.
Connectors never have to upload, install or agree to anything for the map to exist. LinkedIn and mailbox connections are optional enhancers, never prerequisites.
How Boomerang exposes relationship data to agents and people
Boomerang is built to be an input, not another destination. Its agent, Rudy, makes the relationship graph available where work already happens:
- Slack. Rudy answers in deal channels, in the thread you asked in, with a link to the record behind every answer. Rudy for Slack.
- The CRM. Relationship signal as native fields in Salesforce, HubSpot and Attio, so forecasting and scoring agents that read the CRM can read warm coverage too. See the RevOps playbook.
- Any AI assistant, through MCP. Claude or any MCP client can ask Rudy for a warm path. MCP and API access are included on every plan.
The consent model does not change
Feeding relationship data to agents does not mean letting agents spend relationships. Rudy proposes the path. The rep approves the plan. The person who owns the relationship approves the ask and sends it from their own account. Rudy never sends on anyone's behalf. An agent can know that your board member knows the buyer; only your board member decides whether to vouch. That is what keeps the input valuable: a network that gets spammed stops answering.
A quick test for your own AI stack
- Pick three open deals. Ask each of your agents how to reach the economic buyer.
- Note whether any answer names a person in your company, not just a channel or a sequence.
- Ask your executives and CSMs the same question by hand. Count the paths the agents missed.
If the humans find paths the agents did not, the gap is an input gap, and no amount of prompt tuning will close it.
Bottom line
When every sales agent reads the same CRM fields, inbox and call transcripts, they produce similar recommendations. Relationship data is the input that changes what an agent can recommend: from more volume to the one person the buyer will take a call from. Supply it, keep people in charge of sending, and the agents you already pay for start producing warm meetings instead of more noise.
Frequently asked questions
Why do AI sales agents keep recommending cold outreach?
Because cold outreach is the only lever their inputs allow. Most agents read CRM fields, email activity and intent data. None of those record who in your company knows the buyer, so the agent cannot suggest an introduction. It suggests more contacts and more sequences instead, which is a rational answer to incomplete data.
What is relationship data in sales?
It is evidence of who in your ecosystem actually knows a buyer and how well: work overlap, meeting and email history, shared boards, co-authored work, portfolio overlap and similar signals. Boomerang tracks more than 80 of them across executives, investors and advisors, customer champions and partners, and grades each path by strength.
Can I give ChatGPT or Claude access to relationship data?
Yes, for any assistant that supports MCP. Boomerang runs an MCP server, so you can ask Rudy for a warm path to a buyer from inside Claude or another MCP client. MCP and API access are included on every Boomerang plan, and the same data is available in Slack and as CRM fields.
Will an AI agent send introductions on my behalf if it has relationship data?
Not with Boomerang. Rudy proposes the path and drafts the ask, the rep approves the plan, and the person who owns the relationship approves and sends from their own account. Rudy never sends on anyone's behalf. Knowing a relationship exists is different from spending it.
How does Gartner's 10x agents prediction relate to relationship data?
Gartner predicts AI agents will outnumber sellers 10x by 2028, yet fewer than 40% of sellers will say agents improved their productivity. When every agent reads the same inputs, they produce the same cold recommendations. Relationship data is a differentiated input that lets an agent recommend a warm path instead.
How long does it take to add relationship data to my sales stack?
With Boomerang, technical setup takes under an hour and warm paths appear the same day. The graph then keeps itself current, and full go-live typically takes about a week. Connectors do not need to install anything for their paths to appear.
