FBED.

Decision infrastructure for revenue.

FBED builds and runs long-horizon decision infrastructure for the commercial process, delivered as an AI-native service, as software and as decision graphs that measure causal uplift. Using five proprietary models, it plans to produce decision probabilities verified against real-world outcomes at every major decision point. These serve two purposes: training data for commercial AI, and predictable pipeline and insight for revenue owners.

Every decision in the commercial process, recorded and verified by outcome.

Each decision is logged with its state and the probability given to every option, before anything is sent. Outcomes are recorded at every horizon, from a reply within days to a close within months. Every deal becomes one continuous trajectory.

THE INFRASTRUCTURE

Five proprietary models.

Three measure, one decides and one models the buyer's world. Each is a tool that agents call instead of reasoning from raw text, and each must pass its gate on real outcomes before it is used. Every outcome is verified three ways: by what happened, by the person accountable for it, and by our own operators.

All models

What we are building.

Behavioural data in. Probabilistic decisions out.

Six layers, each built on the one before.

Each layer produces the data the next one needs, so none can be skipped.

{{ r.n }} {{ r.name }} {{ r.does }} {{ r.status }} {{ r.when }}

People decide. We record both sides.

A commercial outcome depends on two states: the world (the buyer's readiness and fit) and the person (whether the seller followed the plan). Most systems log neither.

Shannon is built to plan each stage, and to measure how closely the seller follows. The gap between the plan and what happened is where it learns. No environment vendor we've found records it.

{{ s.t }}{{ s.d }}

One record per decision, sealed before the outcome exists.

The action is chosen, and its probability logged, before any language model writes a word. Without the probability, data can only show correlation. With it, results can be tested causally. Whole accounts are randomised, because contacting one person on a buying committee changes how the others respond.

ContextWhich of the ~25 decision points, when, and for which account.
PlanShannon's plan, and how much of it the seller executed.
World stateBuyer readiness at entry and exit, against prediction.
InterventionThe action taken and the probability it was given.
OutcomesEach horizon is verified on its own clock. Pending is never a no.
decision_recordpost-booking · point 9
{
  // context: point 9 of ~25
  decision_point:    "next_stage_after_discovery",
  decided_at:        "2026-10-14T16:02Z",
  account:           "acc_…",  // randomisation unit

  // plan: logged before the call
  plan:              "dec_…",
  seller_compliance: 0.62,  // share executed

  // world state: compressed RWV, % coverage
  world_state:       { entry: 0.54, exit: 0.71 },
  exit_vs_predicted: { rapport: +, trust: +,
                       micro_commitment: −, likelihood: +,
                       problem: +, roi: − },
  expected_decay:    "high without intervention",
  load_bearing:      "in-call commitment to next meeting",

  // intervention: prevent decay
  action:            "hold_commitment_48h",
  propensity:        0.38,
  arm:               "randomised",

  outcomes: [
    { horizon: "next_meeting_held", window: "14d", status: "observed" },
    { horizon: "sql_qualified",     window: "30d", status: "pending" },
    { horizon: "closed_won",        window: "90d", status: "pending" }
  ]
}
THE FUTURE

The reliability of communication between humans and machines.

As physical AI and machines become more embedded in our infrastructure, we believe it is our duty to ensure they are given decisions and tools that support human performance and accountability.

The commercial process, run end to end. Priced on outcomes.

Shannon finds prospects and books meetings through operator-approved machinery. Each meeting arrives with full state and preparation, and your managers and reps work with Shannon through every stage to close.

Meet Shannon.November intake · 15 of 25 sold
Shannon
Command
Find & Meet Data
SRM
Sell
Capacity
Calendar
Connections
SHANNON · THIS MORNING
Nothing needs you today. Two calls Thursday, prep ready an hour before each.

Pipeline is growing and getting warmer, your ceiling has room this month, and every connection is behaving. If any of that changes, it appears here first.

Thu 14:00 · Fenwick Systems · accepted by both sidesprep 13:00
Thu 16:30 · Hollis & Brand · moved by them, from Fridayprep 15:30
PREP · FENWICK SYSTEMS · THU 14:00ready 13:00
WHERE IT STANDSSecond call. Discovery held, both sides accepted Thursday, procurement asked to join.
THE ASKAgree pilot scope, and name who signs.
WATCH FORTimeline pressure reading as price pressure. Hold the number.
View full plan Talk to Shannon
OUTCOME PRICING ONLY
£200per completed meeting

No retainer and no seat fees. You pay only for meetings that happen and pass certification.

THE ECONOMICS
5–10completed meetings per customer won

What a £1,000–£2,000 acquisition budget per customer buys at £200 a meeting. That budget is the level we recommend for a minimum 3:1 LTV to CAC.

01Certified before billingShannon reviews each meeting and issues a certificate of completion and quality.
02Billed after 7 daysMeetings not reported as incomplete within 7 days of the scheduled time are billed.
0330 days to disputeAny certificate can be disputed within 30 days of completion, and is reviewed independently by our research team.

One chain, from first contact to close.

~25 decision points. At booking the work changes: around 1,000 prospects narrow to 8–16 in play, and Shannon moves from acting to planning.

FROM FIRST CONTACT~1,000 prospects
Before bookingShannon acts, at volume.
01FindRanked by how much contact changes the outcome.
02ReachOperator-approved outreach, every move logged.
03BookDelivered with full state and preparation.
FROM MEETING BOOKED8–16 in play
After bookingShannon plans. Your team executes.
04PlanCall preparation, strategy and the next move.
05ExecuteYour rep runs the call. Compliance with the plan is recorded.
06ReflectThe call informs the next stage, or the recovery.
LIVE

Inside meetshannon.ai.

Managers and reps work with Shannon at every stage, in one application.

Planning

The call plan, the strategy and the next move, prepared before each stage.

Execution

Reps work the plan with Shannon alongside, stage by stage.

Reflection

Each call reviewed against the plan. The gap shapes what comes next.

Simulated training

Replay and preplay behavioural simulation before meetings, and manager-set simulations for conversion and time to performance.

SyncingShannon keeps your CRM and calendar current. Every write is logged.
LIVE
Operator approvalOutreach runs only through machinery your operator has approved.
LIVE
ExportersDecision-graph outputs returned into your systems.
IN BUILD

How we prove it.

A randomised share of your target accounts is not contacted. Comparing them with the accounts Shannon works shows the lift Shannon caused, not the lift you would have had anyway. You pay nothing on held-out accounts.

TARGET · DEC 20271.9×
closed deals per 1,000 prospects, against a median outbound team.About 0.84 to about 1.6. Pre-registered, and will be measured against the holdout.
MEASURED SO FAR68%
first meeting held → discovery booked, measured over 18 months.Inside the observation machinery that became meetshannon.ai, since December 2024.

Who it is for.

Sales cycleMulti-stage. No transactional offers.
ChannelOutbound-led go-to-market.
Budget£1,000–£2,000 or more CAC per customer won, for a minimum 3:1 LTV to CAC.

What we ask.

A control groupA randomised share of target accounts held out.
Your stackCRM, calendar and call recordings.
Data rightsAggregate pooling only. Derived use only with consent. Erasure via paulcox@fbed.io.
Live60 ORGANISATIONS
November15/25 SOLD
December25 REMAINING
January50 REMAINING

A decision graph with causal uplift.

Layer 4 estimates the causal effect of each commercial action, by segment and stage, with confidence intervals, from decisions randomised by account. Each edge is a measured effect, not an association.

R&DDecision graphEstimation on logged, randomised decisions.
OPENPartner programmeDesigned with infrastructure partners.
IN BUILDExportersGraph outputs into customer systems.
decision_graph.edgeillustrative · values are not measurements
{
  edge_id:     "edg_…",
  from_state:  "discovery_held",
  action:      "hold_commitment_48h",
  to_state:    "next_meeting_held",
  segment:     { stage: "post_discovery" },
  metric:      "p(next_meeting_held)",

  // effect against the randomised holdout
  uplift:      +0.11,
  ci_95:       [0.04, 0.18],
  unit:        "account",
  estimator:   "doubly_robust",
  refutation:  { placebo: "pass", random_common_cause: "pass" },
  snapshot:    "sha256:…"
}

Where we want partners.

PARTNERWHAT WE BUILD TOGETHERWHAT IT ENABLES
CRM, call-recording and engagement platformsIncremental sync into one canonical model, custom fields included.Complete state at every decision.
Warehouse and database providersImmutable raw storage, a transactional system of record, and a separate analytical store.Replayable history; analysis never runs on the primary database.
Semantic-layer providersOne metric definition shared by people and agents.The same number, to the digit, wherever it is asked.
Key-management providersPer-conversation encryption keys and row-level tenant isolation.Erasure by key destruction.
Causal ML toolingUplift modelling, off-policy evaluation and refutation testing.An effect per edge, with intervals.
Customer systems of recordWrite-back of graph outputs into the tools customers already run.The graph used where the work happens.
Agent platformsAgent access over MCP: governed metrics for AI agents. In development.Agents query the same governed numbers as people.

How we work with systems we connect to.

Source of truth stays yoursWrites are scoped by operator approval, within each vendor's API terms. Every write is logged.
Source-neutralEvery source is an input. We do not compete with the tools we read from.
ReplayableReplaying an ingest adds zero rows. Failures are kept, never dropped.
Written post-conditionsEvery component ships with a written post-condition, checked before release.
What partners get.A co-designed reference architecture for layer 4A joint case study from productionA named integration inside the meetshannon.ai service
Security.Per-conversation encryption keys, row-level tenant isolation, and erasure by key destruction. Security documentation available on request.
Talk to us about layer 4.Partnerships · Sam Bardsley
sambardsley@fbed.io

Verified commercial environments for long-horizon agents.

~25 decision points from prospecting to close. Each decision is drawn from a discrete action set and logged with its compressed world state and the full probability distribution over actions, before any language is rendered. Every task is labelled by the observed outcome.

AVAILABLE NOWQA packsSamples, pass rates and gate reports.
NOV 2026Task setsAll four families, every stage.
DEC 2026 · EXPECTEDRL environmentFull horizon, for QA.
taskillustrative · values are not measurements
{
  task_id:         "tsk_…",
  family:          "off_policy",
  decision_point:  "next_stage_after_discovery",  // 9 of ~25
  horizon:         "weeks",

  // compressed world state, as known at decision time
  state:           { snapshot: "snap_…", world: { entry: 0.54 } },
  actions:         ["hold_commitment_48h", "send_recap",
                     "escalate_sponsor", "wait"],
  propensities:    [0.38, 0.27, 0.21, 0.14],
  arm:             "randomised",
  plan_compliance: 0.62,  // share of plan executed

  // reward from the observed outcome
  reward:          { event: "next_meeting_held", window: "14d" },
  label:           "verified",
  split:           "train",  // test split sealed by account
  snapshot:        "sha256:…"
}

Specification.

FAMILYINPUTSCORED BYHORIZONDELIVERY
Outcome predictionState at decision time and the action taken.The observed outcome.minutes–monthsJSONL + datasheet
Off-policyA proposed policy over logged decisions.Estimators on randomised assignment only.days–monthsJSONL + datasheet
RubricProduce the artefact: research read, opener, call plan.A frozen, human-anchored certifier.minutesJSONL + datasheet
TrajectoryChains of decisions spanning stages, with state between each.Outcomes at every horizon.weeks–monthsJSONL + datasheet
RL environmentAn agent acting at each decision point, prospecting to close.A learned model of buyer and seller behaviour. Fidelity measured against held-out real outcomes by calibration and log-loss on the next outcome, reported in the QA pack.full horizonHosted, Gym-style: reset, step, reward. Dec 2026, expected
EvaluationYour model scored on the sealed test split, with and without FBED's decision layer.The sealed test split.allWith task sets
minutes · same sessiondays · 1–14 daysweeks · 2–8 weeksmonths · up to 3 months

Seller compliance. Post-booking records carry the plan, the seller's measured compliance, and the outcome (intention-to-treat and per-protocol). Close-horizon tasks are released as close attribution is finalised.

Action spaceA discrete action set at each decision point where a policy draws. Language is rendered separately, after the decision, and scored in rubric tasks.
SplitsTraining and validation splits are delivered. The test split is sealed by account and never released; we score models against it on request.
Data assetEach record holds the decision point, the state, the action taken and the outcome, linked into trajectories. It is produced by our own live service, so it cannot be bought or scraped.

The standard every task meets.

Each is a quality gate in the machine. The QA pack reports how many candidates each gate refused, and why.

01LABELS
Real outcomesEvery label is an observed event, never annotation.
AutomaticRecorded by the machine as outcomes arrive.
Censored, not negativeOutcomes not yet verified are censored.
02CAUSALITY
Decision before languageAction chosen and probability logged before any model writes.
Logged propensitiesThe full distribution over actions, at decision time.
Randomised assignmentOff-policy tasks use randomised assignment only.
World state between stagesCompressed world state recorded at every step.
03INTEGRITY
No future leakageState is exactly what was known at decision time.
Sealed held-out setSplit by account, never released, never trained on.
DeduplicatedOne task, one fingerprint.
ReproducibleEvery task traces to a hashed dataset snapshot.
04DIFFICULTY AND RIGHTS
Learnable difficultyFrontier pass rate between 10% and 60%.
Rights-clearedNothing leaves without contractual consent for derived use.

Request a QA pack.

For research, data and partnership teams at frontier labs and environment platforms. Each pack contains:

A task sample with full datasheets: stage, horizon, verifier, label status
Frontier pass rates per model, with attempt counts
Gate report: candidates refused by each gate, and why
Leakage and duplication audit
Held-out split specification and horizon mix
WHAT LEAVES FBED Only derived, de-identified tasks. Never raw transcripts, never component scores. Nothing leaves without contractual consent for derived use. Rubric tasks are de-identified before release (PII redaction, with synthetic replacement of named entities). Method and data-processing terms are provided with the QA pack.
{{ g.label }}
Add your name, work email and organisation.
Request prepared. Your email client has opened with the request addressed to sambardsley@fbed.io. Send it and we'll reply to arrange QA access.

FBED Labs Ltd is the controller of the details you submit. We use them only to respond to this request and arrange QA access, keep them for 12 months, and don't share them. To access or delete them, email privacy@fbed.io.

Two businesses from one machine.

Organisations buy measured lift, priced per completed meeting. Labs buy verified environments. The service earns revenue while it produces the data, and every label comes from an observed outcome.

RAISING{{ raiseAmount }}

Funds the outcome pipeline, the first verified task sets, and growth of the service that produces them.

What we will prove.

A pre-registered target, measured against a randomised holdout: the accounts Shannon works against the accounts it deliberately leaves alone, over the same period and from the same lists.

MEASURED SO FAR68% of first meetings held led to a booked discovery call, over 18 months of observation.
TARGET · DEC 2027 1.9× closed-won new-logo deals per 1,000 targeted outbound prospects, against a median outbound team. Same volume, outbound only. At a 30% outbound share, about +27% of new bookings.
{{ b.s }}{{ b.c }}
{{ b.v }}
Median, closed-won per 1,0000.84
Booked rate is derived; the other stages are published benchmarks. Ranges multiply to between about 0.14 and 1.9. The lift is measured against our holdout, not these benchmarks.

One machine, two businesses.

The service earns revenue while it produces the data. Every label comes from an observed outcome, so the same work serves both markets. FBED Labs turns raw commercial behaviour into probabilistic decisions. Every conversation Shannon runs is logged as context, decision and outcome, building a proprietary behavioural dataset that cannot be bought or scraped. The data advantage is the asset; the service is how it compounds.

ORGANISATIONS
£200per completed meeting

The commercial process, run end to end through meetshannon.ai. Outcome pricing only: no retainer, no seat fees.

FRONTIER LABS Verified long-horizon commercial environments.

Task sets and an RL environment built from real decisions and their outcomes. QA packs available now, task sets in November, the environment expected in December.

Demand for environments.

Labs are paying for training environments. Every vendor in this list replicates software; none records a live market. Deeptune went from a $43M Series A to acquisition by Mercor in four months.

Sources: Epoch AI (Jan 2026); RL List (2026); Dealroom and HeroHunt (2026); New-Tech Europe (Jun 2026); Fortune (Mar 2026).
{{ x.c }} · {{ x.w }}{{ x.r }}

What the round funds.

The outcome pipeline, the first verified task sets, and growth of the service that produces them.

IT HAS WORKED WHEN The pre-registered lift is measured against the holdout, and the first task sets have passed an external lab's quality review.
{{ m.m }}{{ m.p }}

Twelve answers.

{{ q.a }}

The people building FBED.

FBED Labs is an independent research organisation, founded in December 2024 by Josh Bardsley, Sam Bardsley and Abi Firmin. The research, infrastructure and service are built and operated in London.

01 · FOUNDER

Josh Bardsley.

02 · FOUNDER

Sam Bardsley.

03 · FOUNDER

Abi Firmin.

Research, engineering and commercial operation remain connected.

FBED exists to measure commercial decisions across the full trajectory, then use that evidence in the systems that make and execute the next decision.

Contact FBED.