AI agent infrastructure · Consultants & advisory firms

    How consultants and advisory firms use AI agent infrastructure to book 100-400 qualified sales meetings without any technical skills.

    Referrals are wonderful and unpredictable. This is the other half: a system that finds the companies going through the exact thing you fix, while it's still happening.

    Walkthrough video
    Get a quote →Just show me the rules
    The board
    Set up onceOutboundInboundQualify + enrichReach, then decideNo reply → paidConvertAfter the meeting
    Left of the gates, a record is cheap. Right of them, every send spends domain reputation, which you can't buy back quickly. That's why most of the list gets thrown away.
    The conversation is the one step that never gets automated. Everything else on this board exists to put a real person in front of the right person, at the moment they have the problem.
    the loop closes, signed clients rebuild the list
    Rejected (reason logged)
    No verified route. Parked
    Click any box to open itDrag sideways, or open full screen
    What actually happens

    Here's the whole thing, start to finish.

    Eighteen things have to happen between a stranger and a meeting on your calendar. Here they are in the order they happen, so you can see exactly what's involved. Seventeen of them run on their own. The one that doesn't is the conversation, and that's on purpose.

    Build
    1. 01

      Build the agent

      We spin up a server that belongs to you, lock it down, and install the whole toolkit on it. Then we plug it into Telegram so you can just text it. Hermes runs it. So can Claude, or whatever agent you already use. Underneath, ACA holds every lead, sequence, mailbox and conversation.

    2. 02

      Point it at the sources and build the list

      No bought database. We point it at the places where companies show they've got the problem: job posts, funding news, maps, review sites, communities, your competitors' followers. Then we leave it running, so the list keeps growing.

    3. 03

      Filter the leads

      Five gates, cheapest first: hard rules, then how fresh the signal is, then an ICP score, then exclusions, then suppression. Every rejection gets a reason written next to it.

    4. 04

      Enrich and verify

      We find the actual decision maker and a way to reach them that works. Several providers in a row, then we verify it ourselves. If it comes back invalid, we go looking again instead of sending anyway.

    5. 05

      Build the outbound engine

      Separate sending domains, set up properly, and every mailbox warmed before it sends a single campaign email.

      This one isn't sequential. Warmup takes weeks, so it starts on day one. Otherwise you finish the list and then sit there for a month.

    6. 06

      Build ten sequences and the LinkedIn campaigns

      Ten sequences. Several offers, several openings, all written to be compared. You don't pick the winner here. The market picks it, usually inside a few weeks.

    Launch
    1. 07

      Route the leads into them

      We load the verified list in and split it so the offers get tested against comparable segments. Suppression gets re-checked at send, not at build.

    2. 08

      Build the inbound engine

      Lead magnets on LinkedIn, deliberately narrow, so the wrong person doesn't bother asking for one.

    3. 09

      Automate delivery, filtering and the first line

      It captures the request, delivers instantly, scores them against the same gates the outbound list gets, and opens with a real question instead of a thank-you page.

    4. 10

      Hash everyone enriched into ad audiences

      Every contact we enrich gets turned into a fingerprint before it leaves your server. Meta and Google can match it. Neither of them ever sees your list.

      Also not sequential. This one fills up every time step 04 finds someone.

    5. 11

      Launch the campaigns and the page they land on

      LinkedIn, Meta and Google, all pointed at one page. Not a brochure: a calculator that runs their numbers, with the booking form right there on the same screen.

    Convert
    1. 12

      Monitor replies manually

      A person reads every reply, from both lanes. The second someone replies, every sequence they're in stops. Everywhere. The agent has already pulled the thread, the signal and the score together, so nobody starts cold, and it drafts a reply for the person to change or bin.

    2. 13

      Book the meeting

      On the calendar, with a brief attached: what the signal was, what they scored, and the whole thread.

    3. 14

      Pass the booking back to the agent and the ad platforms

      A booking is a conversion. It goes back to the agent so it learns who books, and to Meta and Google so they stop optimising for clicks and start optimising for meetings.

      Happens the same second as 15. Both fire off the booking.

    4. 15

      Send the pre-appointment sequence

      Testimonials and case studies that match their situation, plus the confirmation and the no-show nudges. The gap between booking and the call is where most meetings quietly die.

    5. 16

      Send the summary and the proposal

      Same day, while it's still warm. A recap in their words, and a scope built from what they actually said.

    6. 17

      Run the post-appointment follow-up

      Paced, not nagging, with material aimed at whatever they pushed back on. It runs until you get a yes or a clean no.

    Compound
    1. 18

      Signed, and the loop tightens

      The win goes back to Meta and Google as a lookalike seed, and back to the agent so it rewrites the ICP from who actually bought. Offers that lost get retired. Round three beats round one because it's built on results instead of guesses.

      Different event from 14. Booked teaches it who books. Signed teaches it who buys.

    Why an agent

    It does the 90% that isn't selling.

    The hours in outbound aren't in the sending. They're in the research: reading a company's site to work out whether they're a fit, hunting down the right person, checking you haven't already emailed them, writing an opener that references something real. Done properly that's ten or fifteen minutes a prospect. An agent does it in seconds and shows its working. Across a few thousand companies, that's a full-time job you get back.

    And it isn't smarter than your best salesperson. It's more consistent. It doesn't get bored on row 400, doesn't skip the suppression check on a Friday, and never sends without verifying first.

    The agents handle this

    High volume, low judgement.

    • Watch every source, all day, and catch the trigger the hour it appears
    • Read a company and score the fit, with the reasoning written down
    • Find the right person and a route that actually resolves
    • Check exclusions and suppression before every single send
    • Draft the first line from the signal that surfaced them
    • Pick the mailbox, pace the send, watch inbox placement
    • Sort replies by intent and stop every sequence the moment one lands
    Only a person does this

    Low volume, high judgement. The agent drafts and briefs. The person decides and sends.

    • The conversation. All of it, assisted by AI
    • Deciding whether a lukewarm reply is worth pushing on
    • Setting the offer, the positioning and the price
    • The meeting, the proposal, the negotiation
    • Judging whether an odd-shaped company is worth an exception
    • Anything where being wrong is expensive

    That split is the whole design, not a limitation we're working around. The agents take the part that's volume so a person has time for the part that's judgement. Take the agents away and someone does ten thousand research tasks by hand. Take the person away and you have a robot negotiating with your next client. Neither of those is a business.

    Here's the crew, and what each one is allowed to touch.

    Scout

    Finds the signal

    Watches every source continuously and picks up the moment a company shows it has the problem. Merges duplicates into one record with every reason attached.

    source watchers · scrapers · dedupeFast model + rules
    Analyst

    Decides if they fit

    Reads the company like a good salesperson would: the site, the positioning, who they sell to. Scores it against your ICP and writes down why.

    company read · ICP scoringFrontier model
    Warden

    Says no

    Exclusions, suppression, existing customers, anyone who asked not to be contacted, anyone who bounced. Runs on rules. No model gets a vote.

    suppression · exclusions · DNCRules only
    Resolver

    Finds the route

    Runs the providers one by one, verifies the result itself, and goes looking again when something comes back invalid instead of sending anyway.

    providers · verifier · re-searchRules + fast model
    Scribe

    Writes the opener

    Builds the first line from the signal that surfaced this company, in your voice. Then checks its own work for broken variables before anything ships.

    copy draft · variable checkFrontier model
    Dispatch

    Gets it delivered

    Picks the mailbox, respects the daily cap, paces it like a person, rotates domains, and pulls one out the moment placement drops.

    mailbox rotation · pacing · placementRules only
    Listener

    Reads the reply

    Reads every reply and sorts it: interested, out of office, wrong person, bounce, not now. Stops every sequence that lead is in, then wakes a human with the full thread.

    intent classify · sequence stop · notifyFrontier model
    Curator

    Closes the loop

    After a deal signs, works out what the buyers had in common, rewrites the ICP filters, and seeds the lookalikes for the next round.

    trait analysis · audience seedFrontier model
    Buying signals

    What we watch for in your market.

    Anyone can pull a list. The difference is timing: four moments where the same message goes from ignorable to obvious.

    Trigger 01

    A leadership change

    New executives rebuild, and rebuilding is when outside help gets hired, usually in the first two quarters.

    Trigger 02

    A funding round, merger, or acquisition

    Integration work always follows, and internal capacity almost never does.

    Trigger 03

    They announced a transformation programme

    A public commitment with a deadline attached is the most reliable trigger in professional services.

    Trigger 04

    They're hiring for the problem you solve

    They've already decided to spend. The only open question is whether it's a hire or you.

    The short version

    13 things we learned the hard way.

    If you take nothing else from this page, take these. Most of them you can use tomorrow, with or without an agent running them.

    1. 01
      Never send cold volume from your primary domain.It's the address your contracts come from. Buy separate ones that read like you, and keep the real one out of the blast radius.
    2. 02
      A new mailbox sends nothing for weeks.Warm it, then ramp. A mailbox that jumps straight to volume lands in spam and tends to stay there.
    3. 03
      Many mailboxes at small daily caps, never one at a big one.One mailbox sending a thousand messages is the clearest spammer signal there is.
    4. 04
      If you can't name the trigger in the first line, don't send it."You might need this" and "you were dealing with this last Tuesday" are not the same email.
    5. 05
      A signal older than a quarter is trivia, not a trigger.Freshness is a filter, not a nice-to-have. Old signals read as research, not relevance.
    6. 06
      Rules for facts, models only for judgement.No model should get a vote on whether a company has eleven employees. Save the expensive thinking for whether they actually fit.
    7. 07
      Log the reason for every rejection.A funnel you can't audit is a funnel you can't fix. You want to know why 9,000 rows got dropped.
    8. 08
      Verify every address, and re-search the invalid ones.A pattern-guessed address bounces, and bounces are charged to the domain, not to the row.
    9. 09
      Check suppression before every send, not once at list build.People become customers, complain, or bounce between the build and the send. Most leaks happen in that gap.
    10. 10
      Measure inbox placement, not open rates.You can post a healthy open rate and still be in the spam folder for the segment that actually matters.
    11. 11
      A reply stops every sequence that person is in, immediately.Nothing kills a warm conversation faster than an automated follow-up two days after they said yes.
    12. 12
      No reply is not a no. Retarget instead.Most people never reply to a first email. Put them in a hashed ad audience and let recognition do the work the inbox couldn't.
    13. 13
      Let whoever signs rewrite the ICP.Build the next list from the traits of people who actually bought, not from the profile you guessed at the start.
    FAQ

    Common questions.

    Is this just an AI that sends emails?+
    No. The sending is the smallest part. Most of the work is upstream, finding companies showing they have the problem right now, throwing away the ones that don't fit, finding a contact route that actually resolves, and building sending infrastructure that gets messages delivered. And it stops entirely at the reply: a person handles every real conversation.
    Where does Hermes actually run?+
    On a VPS provisioned for you, not inside a shared app you rent by the seat. We secure it first (key-based access, firewalled, isolated, backed up), then install its skills, dependencies and agents, then connect it to Telegram so you can talk to it from your phone. It stays on permanently, which is the point: the source watchers keep running overnight and at weekends, which is when a lot of the signals actually appear. Your keys and your prospect data stay on your machine.
    Why build secondary domains instead of just using ours?+
    Because cold volume from your primary domain puts the one address you can't afford to lose at risk: the one your invoices, contracts and client threads come from. Secondary domains are registered, authenticated and warmed to look and read like you, and they carry all the cold volume. If one gets damaged, it is replaced. Your real domain is never in the blast radius.
    What actually happens to our contact data before it goes to ad platforms?+
    It gets turned into a fingerprint before it leaves. The platform can match that against its own users, not a readable list of your prospects. Customers and existing opportunities are excluded from acquisition audiences and used only as lookalike seeds.
    How long before meetings start appearing?+
    The list build and qualification is the fast part. The constraint is infrastructure: mailboxes have to warm before they can carry volume, and rushing that is how campaigns land in spam permanently. Expect the first weeks to be building and testing, with volume ramping as reputation earns it. Anyone promising meetings in week one is skipping the warmup.
    Which AI model do you use?+
    Several, deliberately. Deterministic rules handle facts like headcount and geography, no model gets a vote on those. Fast, cheap models do high-volume sorting. The strongest models are reserved for genuine judgement calls, like reading a company and deciding whether it truly fits. The system is model-agnostic and runs on your own keys, so it improves as the models do without you being locked to one vendor.
    What if our ICP is wrong?+
    It usually is, at first, everyone's is. That is what the loop is for. The traits shared by people who actually signed become filters for the next build, and the traits shared by people who ghosted get filtered out. Round three targets meaningfully better than round one because it was built from outcomes rather than assumptions.

    The same playbook, for your industry.

    Get started

    Point this at your market.

    Tell us who your practice sells to. Hermes builds the list, the filters run, the infrastructure warms, and the first conversations start, with a human on your side of every one of them.

    Get a quote →See pricing

    Your model, your keys, your tenant. Cancel anytime.