Briefing

AI, plainly —
and what it is worth
to this business

Twenty minutes. First, enough of the technology that we all use the same words. Then the numbers for a data management practice of this size.

Clinical data management01

Part one — the technology

A language model does one thing

Given a sequence of text, it predicts what comes next. Everything else is built on top of that.

your text
tokens
next-token probabilities
chosen token

The loop runs once per token. What people call "reasoning" is the model spending more tokens thinking before it answers.

BackgroundKarpathy, Intro to LLMs youtu.be/zjkBMFhNj_g
The prediction loop02

Part one — the technology

Four words worth knowing

Token

A chunk of text, roughly four characters. Models read, price and limit work in tokens, not words.

Context window

How much text it can hold at once. Everything it knows about your task lives there, and is gone on the next call.

Weights

The trained parameters. Open-weight models can be downloaded and run on your own hardware.

Temperature

How adventurously it picks. Low for data work and code. Higher for drafting and ideas.

ReferenceHugging Face, Gen AI Course — tokenizers & concepts huggingface.co/learn/llm-course
Vocabulary03

Part one — the technology

What it does well, and where it breaks

Reliable

  • Transforming text you supply
  • Structure: summaries, tables, mappings
  • Writing and explaining code and scripts
  • Language, tone, translation
  • Following an explicit written procedure

Unreliable

  • Facts it was never given — it fills gaps confidently
  • Arithmetic and precise counting
  • Anything after its training cutoff
  • Your files, your systems, your live study data
  • Remembering yesterday

The right-hand column is the specification for what an agent has to solve.

EvidenceDell'Acqua et al., Jagged Frontier (BCG·HBS·MIT; 700+ consultants): +near-40% inside, −19pts outside SSRN 4573321
Capability boundary04

Part one — the technology

A chat window gives the model words.
An agent gives it hands, memory, and a job.

The shift05

Part one — the technology

Hermes is the loop around the model

You send a request. Hermes gives the model tools, reads the real results back, and repeats until the work is actually done.

request
model decides
tool runs
real output
model reads it

It never guesses what a script returned. It runs the script and reads the result.

DocsNous Research, Hermes Agent Quickstart — the agent loop hermes-agent.nousresearch.com
The agent loop06

Part one — the technology

Four things Hermes adds

01

Tools

Files, spreadsheets, databases, scripts, web, email. How the model touches anything real.

02

Memory

Durable notes loaded every session, so what you explained once does not need explaining again.

03

Skills

Your procedures on disk, loaded only when the task matches. Your SOPs, not the model's defaults.

04

Working unattended

Scheduled and background work, so reconciliation, listings and checks run overnight and are waiting in the morning.

DocsNous Research, Hermes Agent — tools, memory, skills, autonomy hermes-agent.nousresearch.com/docs
What the framework contributes07

Part one — the technology

Skills are where your expertise lives

A skill is a written procedure with a trigger. Hermes loads it only when the work calls for it.

  • The exact steps, not a paraphrase
  • The pitfalls you have already paid for
  • Your standards for that kind of work
  • Reusable by everyone, on every study

Why this matters here

Your value is twenty years of knowing how a sponsor wants their data. A skill is that knowledge written down once, applied consistently, and no longer trapped in one person's head or lost when they are on leave.

DocsNous Research, Hermes Skills — procedures on disk hermes-agent.nousresearch.com
Procedural memory08

Part one — the technology

Protea — a privacy and memory layer

Protea is a privacy and memory layer, used as the example of the right architecture for regulated work. It sits between a client and any model backend, so you get the frontier model and keep the data.

Redaction — the killer feature

Named entities — people, sites, sponsors, compounds — are replaced with stable placeholders before the request leaves your hardware, and restored in the reply. The model never sees a real patient name or a blinded molecule, so you can run frontier models without an exception.

Local, one-file memory

A single SQLite file per user behind an OpenAI-compatible API. An embedded local model means it runs fully offline once the model weights are pulled. No third party ever holds your corpus.

It remembers

Structured recall across conversations and months — spiral-indexed by relevance and recency — rather than starting from nothing every session. Fifty years of knowing how a sponsor wants data stops living in one person's head.

And it gets your work done

Conductor routing passes each task to the model best suited for it; a confidence cascade escalates only when needed: local model for the routine, frontier model for the hard call. Consensus fans a delicate question to several models and synthesises the answer. Teams decompose a brief, rate deliverables, and review every result. An Open WebUI front-end gives clean per-user separation, with identity forwarded through the API.

The pitch to a sponsor: "we run AI-assisted data management and your data never leaves our environment." Very few small vendors can say that sentence truthfully.

SourceProtea docs/README.md; redaction pipeline docs/redaction-algorithm-research.md (local: D:\Projects\Protea\docs)
Protea — SciBit09

Part one — the technology

The failure mode to design around

Asked for something it cannot know, a model produces a fluent, plausible, wrong answer rather than nothing.

How it shows up

An invented variable name. A confident number with no source. A mapping that looks right and is not.

What contains it

Read before answering, run before claiming, show the source. Grounding in real output — and a person signing off. Never autonomy over a regulated deliverable.

EvidenceNavigating the Jagged Frontier — confidence without calibration HBS 64700
Hallucination10

Part one — the technology

The experienced person is the one who makes AI safe

The evidence is not "experts get faster" — it is that expertise is what lets AI be used on the hard stuff at all.

Knows what to ask

The model cannot produce the question that only twenty years of knowing sponsors produces. The value is upstream of the answer.

Catch the wrong answer

AI is fluent and confidently wrong. Inside its capability boundary it gains ~40%; pushed past it, performance drops 19 points. Only experience says where that boundary is.

Ground it in reality

The model suggests; the domain expert decides. A 19-point penalty exists precisely because it is not obvious to a generalist which tasks AI can do and which it cannot.

Turn it into procedure

A senior hand encodes their own judgement as a skill. Their expertise stops being a bottleneck and becomes repeatable.

Why a senior person wins

AI compresses the skill gap on routine tasks — a beginner with AI reaches near-expert output on well-defined work. But a data management practice does not bill for routine. It bills for the messy, consequential, one-off judgement where a confident wrong answer costs a study.

The leverage is trust, not speed

The real gain for an expert is not doing the same task faster. It is having a tireless assistant who drafts and cross-checks, while the expert spends the saved hours on the calls, the protocols, and the relationships no model can hold.

EvidenceOECD, Unlocking productivity with generative AI (2025): experienced workers gain when AI complements their expertise OECD 2025
Evidence: OECD (2025), Jagged Frontier study (BCG / HBS / MIT, 700+ consultants)11

Part two — the numbers

The question is not whether AI is impressive.
It is what an hour of your team's time is worth, and how many of those hours you currently sell.

Shifting to the business12

Part two — the numbers

Where the hours actually go

A worked example at the scale you described: eight people, about 1,700 working hours each — roughly 13,600 hours of capacity a year. Adjust the staff count and the numbers scale.

Billed to sponsors

9,520 hrs

70% utilisation. This is the revenue.

Not billed to anyone

4,080 hrs

Status reports, proposals, reconciliation chasing, formatting, re-keying, QC of your own paperwork, answering the same sponsor question a fourth time.

That second box is the target. Not the people — the 4,080 hours of work nobody pays for.

RateContract CRA rates $85–110/hr; CRA salaries $95–115k (market reporting 2025–26) — defines revenue per FTE
Capacity, not headcount13

Part two — the numbers

What the industry pays per query

Cost to resolve one query

$28–71

Published benchmark range. Medidata's analysis puts the average at $53.87.

Queries in an average Phase III

~96,980

Which is roughly $5.2m of query handling in a single large study.

If a quarter never get raised

~$1.3m

Avoided cost, on one study. Cleaner data at entry beats faster query resolution.

You do not have to capture that whole number. You have to be the vendor who can credibly claim a share of it and price accordingly.

CostMedidata: avg $53.87/query; ~96,980 queries per avg Phase III (industry reporting)
Sources on the final slide14

Part two — the numbers

Three paths, same business

Revenue, USD. Modelled on 8 billable staff at ~$1.0m in 2026, headcount growing 4% a year in all three cases.

ModelAdoption ceiling anchored on Tufts CSDD/DIA (Getz, Applied Clinical Trials, Jun 2025): adopters reported ~18% average cycle-time reduction, but only 11% of companies had fully implemented — not a vendor 35–45% projection
Assumptions on slide 2115

Part two — the numbers

The gap is in the profit, not the revenue

EBITDA, USD. Waiting is not neutral: from 2028 the client-side expectation of AI-era pricing arrives whether or not you have the capability.

ModelMargins on ~$1.0m base; from 2028 non-adopters face ~3.5%/yr real rate erosion (see slide 21)
Same three paths16

Part two — the numbers

2031, side by side

PathUtilisationRevenueEBITDAMarginCumulative EBITDA 2026–31
Start now (2026)82%$1.62m$481k29.7%$2.14m
Wait until 202980%$1.52m$383k25.2%$1.34m
Do not adopt70%$1.16m$57k4.9%$0.80m

The cost of a three-year wait

$802k

Cumulative EBITDA foregone by starting in 2029 instead of 2026. The capability is still reachable later. That money is not.

What it costs to start

~$28k

First-year programme: tooling, hardware, and the time to write your procedures down properly. Roughly 2% of revenue thereafter.

ModelSame assumptions as prior slides; see slide 21 for full parameter list
Summary17

Part two — the numbers

Read the model carefully

Every scenario on the previous slides has the team growing, not shrinking. The gains come from selling more of the hours you already have, not from having fewer people.

What changes

The dull half of the job. Re-keying, chasing, formatting, first-draft documentation, the fourth version of the same listing.

What does not

Judgement on a protocol. The sponsor relationship. Who signs off. Knowing which query actually matters and which is noise.

What grows

The scope one data manager can carry. More studies each, more senior work each, and a reason to pay them more.

A practice this size does not win by employing fewer people. It wins by being able to take on the study it would previously have turned down.

EvidenceOECD (2025) & Stanford (2024): AI raises output without cutting headcount Stanford SIEPR
People18

Part three — where to point it

Inside the work you already do

Protocol → build

Draft the eCRF spec, edit checks and validation rules straight from the protocol, for a human to correct rather than compose. The blank-page hours are the expensive ones.

Query triage

Rank the day's queries by whether they will actually change an analysis. Draft the text. The data manager approves or rewrites.

Reconciliation

External labs, SAE against the safety database, vendor files. Overnight, with an exceptions list on the desk each morning.

Mapping to standards

First-pass SDTM mapping and define documentation. Tedious, rule-based, and reviewable — the ideal shape for this.

Coding support

MedDRA and WHODrug first-pass suggestions with the reasoning shown, and a person confirming every term.

The paperwork layer

DMP drafts, status reports, meeting minutes, sponsor updates. Pure non-billable time today.

Proposals

Bid responses and costing drafts from your own past work. More bids answered, from the same people.

Onboarding

A new hire asks the system how you do things and gets your actual procedure, not a guess. Weeks off ramp-up.

PracticeA scoping review of AI in clinical trial risk assessment — 142 studies (2013–2024), npj Digital Medicine Nature · PMC — AI & the future of clinical trials
Efficiency inside the current model19

Part three — where to point it

Work you cannot sell today

The interesting part is not doing the same work cheaper. It is that some services only become viable once the marginal hour is cheap.

Continuous data surveillance

Not a monthly cleaning pass — a standing service watching every incoming record. Recurring monthly fee instead of project fees.

Rescue and remediation

Inherited messy databases from failed vendors. Currently priced out of reach because it is all manual. High-margin, and it wins the next study too.

Legacy data conversion

Old studies into modern standards for regulatory submission or reuse. A large backlog exists and almost nobody wants the work.

The small-sponsor market

Biotechs and academic units the big CROs will not quote for. A cost base this low makes small studies profitable.

On-premise as the pitch

Sponsors nervous about AI and confidentiality. Local models mean you can offer the benefit without the data leaving. Protea is exactly this.

Africa as the position

Trials on the continent, with local regulatory knowledge, local time zone and a cost base that still undercuts Europe. A hard-currency price point.

MarketAI for clinical trial design, conduct & analysis — npj review PMC 13040932 · site-selection/enrollment study PMC 10927105
New scope, new markets20

Part three — where to point it

What the first ninety days look like

01

Pick the single most repetitive non-billable task. Time it honestly for two weeks first, so the before number is real.

02

Write that procedure down properly as a skill. This is the actual work, and it has value even if the AI part fails.

03

Run it alongside the human for a month. Compare outputs. Keep the human sign-off permanently.

04

Measure the hours recovered. If they are real, take on one more study with the same team. That is the whole thesis.

What the model assumes

8 billable staff, 1,700 hours each, 70% utilisation and a $105 blended hourly rate — giving ~$1.0m in 2026. Headcount grows 4% a year in every path. Utilisation rises to 82% at full adoption, phased over roughly four years. Programme cost ~$28k in year one, then ~2% of revenue.

From 2028, non-adopters face ~3.5% annual real rate erosion as sponsors price in AI-era expectations. This is the single most debatable assumption and the main driver of the "do not adopt" line.

Sources: contract CRA rates $85–110/hr and CRA salaries $95–115k (market reporting, 2025–26). Query cost $28–71, average $53.87 (Medidata PICAS/CROCAS analysis); ~96,980 queries per average Phase III (industry reporting). Adoption effect anchored on Tufts CSDD/DIA (Getz, Applied Clinical Trials, Jun 2025): adopters reported ~18% average cycle-time reduction; only 11% of companies fully implemented. Not a vendor 35–45% projection.

SourcesRates, query cost, and productivity ceiling as on slides 13–15. Full parameter list & citations here.
Start small, measure honestly21

Close

The technology is ordinary. It predicts text, and a framework gives it tools and memory.

The opportunity is not ordinary. Four thousand unpaid hours a year, in a business that bills by the hour.

Nothing here requires betting the company. It requires writing down how you work, and starting.

End22
1 / 21  ·  ← → SPACE  ·  F FULLSCREEN