Fin resolves 76% of tickets, so Intercom stopped counting them in staffing math

The last two years were about agents writing copy and answering easy tickets. The story moved. Now agents do the work your team used to grind through by hand: researching prospects, baiting scammers, forecasting staffing. The software layer underneath your operation is getting rebuilt around them. Let me catch you up on where agents actually earn their cost, and where they quietly add new risk.

The deep cut

• Agents earn their cost on grind work, not judgment. Floqer cut prospecting time 70% by automating research, not decisions.
• Measure the agent before you staff the humans. Fin resolves 76% of queries, so Intercom forecasts only human-bound volume.
• An old acknowledgment is not a current promise. Alice's accepted handoff kept telling customers she owned a request Bob already took.

Where the hours actually go

The clearest wins come from jobs that eat time but need no taste. Prospect research is one. Floqer runs a fleet of agents that search across 100-plus data sources to build target lists, verify contacts, and update records when someone changes jobs. Customers report cutting prospecting time by as much as 70%. The company passed $1.2 million in ARR and grew revenue roughly tenfold in a year.

Fraud intelligence is another. Apate sends agents undercover to get scammed on purpose, wasting scammers' time while pulling out bank accounts and crypto wallets. Commonwealth Bank saw a 76% drop in publicly reported scam losses. The pattern holds: pick a task that is repetitive, measurable, and currently manual. That is where agents pay for themselves.

Plan for the team you actually have

Agents do not just add capacity. They change what your humans handle, and that breaks old planning math. Intercom's new Workforce Management makes the point plainly: its agent Fin resolves 76% of queries on average, and that number climbs weekly for many teams.

So forecasting total ticket volume no longer tells you how many people to staff. Intercom now builds forecasts that exclude everything Fin handles, so you plan around human-bound work only. That is the real shift for a leader. Your staffing model needs to isolate what lands with humans, or you will over-hire for volume an agent already absorbed. If you still run capacity planning in a spreadsheet that counts every ticket, you are planning for a team that no longer exists.

The new front door is a coding agent

If you sell to developers, your buyer changed. When a developer tells Claude Code "add a database," the database the agent picks is the one that ships. One write-up from talking to developer-tool founders found no single "agent recommendation," with one dataset of 1,650 prompts across nine agents showing different tools win on similar tasks.

Being known is only half of it. Supabase's own post, cited in that piece, admits agents know about Supabase but do not always use it right. Their fix was a 100-line SKILL.md that teaches agents how to think about the product, not just what it does. For your team: typed SDKs, clean docs at the paths agents expect, and an llms.txt file now function as distribution, not polish.

The handoff that lies

Here is the risk that will not show up in a demo. A synthetic walkthrough of a support handoff shows Alice accepting a request, the request moving to Bob, and the assistant still telling the customer "Alice is your assigned support contact." Every event in that acknowledgment is real. The message is still wrong.

The bug is treating an old acceptance as a permanent promise. Once a request moves, "Alice accepted at 10:04" is history, but "Alice is your assigned contact" is a claim about right now. As agents chain work together, these stale states multiply. The lesson for your roadmap: decide when responsibility starts and ends, and make the agent check the current state before it speaks, not the last state it remembers.

Three questions for your team

• Which repetitive, measurable task on our team would an agent own first, and what number proves it worked?
• Does our staffing model forecast human-bound volume separately from what agents resolve, or are we still counting every ticket?
• Where in our agent workflows can a stale state turn into a confident, wrong promise to a customer, and who owns catching it?