The head of CMC who had already left
A contract drug manufacturer is hiring an inside sales rep to research every target account and write a brief before anyone reaches out. Here is how a morning of briefs goes stale, what went wrong for others who handed the research to a machine, and a build that holds up.
How to use this page
Read the story, then the map. If no one in your business holds the middle square, start there, not with the build. If someone does, read what went wrong for others, then copy the build kit into your AI agent.
On 14 October 2026 a contract development and manufacturing organization for oral solid dose drugs (tablets and capsules), about 60 people and two sites in Southern California, posted for an inside sales representative. The lines that carry the job: “Research accounts to understand each prospect's product, development stage, timeline, decision makers, and potential manufacturing needs before outreach”, “Prepare account briefs and prospect research to support business development conversations and follow-up”, and “Maintain consistent follow-up with prospects throughout long and complex pharmaceutical sales cycles.” Pay: “$72,000-80,000 annually”, plus an incentive plan.
Here is how it goes today. Monday at 8:30 the rep opens the week's list of 20 target companies and starts on a biotech with an oral drug entering Phase 2, the mid-stage trial that needs real batches made. Website, pipeline page, the public trial registry, press releases, professional profiles, one tab at a time. The brief is done by 9:35: the product, the stage, a timeline, and the head of CMC (chemistry, manufacturing and controls: the person who picks a manufacturer) named on the company's site. Six briefs by lunch. No email sent.
On Wednesday the rep emails the head of CMC from the brief. It bounces. He left in August; the site was never updated. The new one is choosing an oral solid dose partner this month, and the trial start that would have told her so was announced three weeks ago on a page the rep did not open. On day 20 the biotech signs its Phase 2 drug product work with another manufacturer. A development program with clinical batches runs about $100,000 to $400,000 on one marketplace's 2026 figures.
Nobody here is careless. The posting asks for exactly this research, and the rep did it properly, at the pace people who do it report: an outbound agency measured “3+ hours per 10 accounts when done properly”. What a hand-written brief cannot do is stay current. The facts that decide the timing, a trial start, a financing, a new head of CMC, are the ones that move, and a page read on Monday says nothing about what changed on Tuesday. The time comes out of the outreach the brief was for.
Hand the whole thing to a tool that researches, writes and sends, and you get the other failure: the fabricated funding round, the wrong title, the “30% cringe” that one agency's reviewer says is “enough to tank your credibility”. The build below keeps the reading with the machine and the email with the rep.
The biotech, the head of CMC, the days and the counts are an example built from the posting, which gives no counts. The 30 minutes a brief, the 20 briefs a week, the $17 and the $100,000 are estimates from public posts and one marketplace's price range.
The map: the same job as a system
The judgment it all runs on sits in the middle square, held by a person. Around it, the work in the order it moves, each square run by the machine or by a person. Along the bottom, what it runs on.
The middle is which account is worth a call, and what to say: whether the company can make this program, who there decides, and the angle. The rep holds it, with the VP of business development.
The same job as a system. The middle is the judgment about which account is worth a call and what to say, held by the rep. The target list comes in on the left; across the top the machine gathers the facts, spots the trigger and drafts the brief; on the right the rep writes the email. Along the bottom is what it runs on: the schedules and the switch, the counts, and the profile and fields that tell it what to look for.
- Runs on its own
- A person does it
- The judgment it all depends on
- What it runs on
Who holds the middle square
The inside sales rep, with the VP of business development. Which pharmaceutical or biotech program the company can actually make, who there decides on manufacturing, and what the first email should say to that person.
Before you build, these must already be true in your business:
- The company has written down which products and stages it takes on, so the list is built on that and not on a database filter.
- The target list lives in one CRM, one record per company, each with its website and an owner.
- The rep reads every brief before using it and writes the first email in the rep's own words, under the rep's own name.
Still deciding whether to do this, or no one holds this judgment yet? See the verdict: what it costs today, what to hand over, the bill and the owner
What went wrong for others, and the rules it gives the build
Five things that went wrong for people who did this job, or that tripped them up, and the rule each one gives the build.
“AI generated first lines are 70% good and 30% cringe. that 30% is enough to tank your credibility with prospects who can smell automation instantly. 15 minutes of human review per 50 accounts is cheap insurance”; the agency's answer: “the QA part is non-negotiable”.
A reply to an outbound agency's own workflow post, and the agency, April 2026. Read it
The rule: The machine drafts the brief only; the rep reads every brief and writes the email in the rep's own words.
“credits do add up if you're not careful. we try to control it by only enriching after signal filtering and avoiding unnecessary enrichments upfront”; and “it requires an expert to handle and utilize the credits”.
The same agency, April 2026. Read it
The rule: Research only the companies on the rep's list, never the whole market, and count the cost per brief every week.
“if the data layer is messy no amount of researching is going to fix it”.
A revenue operations lead, July 2026. Read it
The rule: Clean the target list in the CRM before switch-on: one record per company, with its website and an owner.
“The agent generates ‘personalized’ outreach that fabricates a fact”: “A non-existent funding round”, “A misattributed quote”, “A wrong title”.
A sales coaching firm's review of AI sales agents, June 2026. Read it
The rule: Every line of the brief carries its link and the date it was read; a line with no link is deleted, and the count of deleted lines is an alert.
“Send without review: the agent sends emails without review, based on your outreach settings.”
The CRM maker's own help page, October 2026. Read it
The rule: No send mode, ever: every draft waits in the rep's queue, and the machine has no email account.
What it costs to run
About $0.21 a brief in usage (10 searches at $7 a thousand, 15 pages at $1 a thousand, about 50,000 tokens in at $2 a million and 3,000 out at $10 a million); about $17 a month for 80 briefs, about $21 a month for the daily check (100 accounts, one search a day, 30 days, at $0.007 a search), plus n8n Starter (20 euros a month, 2,500 workflow runs; this uses about 110: 80 briefs and 30 daily checks). Rep time: five minutes a brief, about $2.90 at $34.62 an hour, against about $17 by hand.
Estimate. The token counts, the five minutes a brief, the 80 briefs a month and the 100 watched accounts are assumed. The timed week replaces them.
The build kit
One prompt to paste into Claude, ChatGPT or any AI agent that can build. It checks the middle square before anything else, builds the map square by square, carries the 5 rules above, and has the whole build: what starts it, every step, the data, the review, the alerts, what to do when it breaks, and what to count before you switch it on.
Before you build
- It may never, on its own: Send or schedule an email; guess an email address, a timeline or a decision maker; write a line without a link; change a CRM field; or read or store anything a prospect sent us in confidence.
- Keep with a person: Which accounts go on the list, because only a person knows what the company can make; the angle; the first email, in the rep's own words under her own name; the call; which line to trust when two sources disagree; and anything a prospect has said in confidence.
- Not yet: Sending any email on its own; guessing an email address, a timeline or a decision maker the sources do not state; scoring an account as ready; writing to any CRM field other than attaching the brief as a note.
Help me build this in my business: Researching each target account before anyone reaches out, and writing the brief: the prospect's product, its development stage and timeline, who decides on manufacturing, and what it is likely to need. Then the outreach, the first meeting for the business development team, and the follow-up through "long and complex pharmaceutical sales cycles". Before you build anything: this runs on one person's judgment, "Which account is worth a call, and what to say", held by the inside sales rep, with the VP of business development. Ask me whether each of these is true in my business, one at a time: 1. The company has written down which products and stages it takes on, so the list is built on that and not on a database filter. 2. The target list lives in one CRM, one record per company, each with its website and an owner. 3. The rep reads every brief before using it and writes the first email in the rep's own words, under the rep's own name. If any is not true, stop and tell me which, and what I need in place first. Do not build around it. Rules from others who built this and got it wrong. Keep every one: - The machine drafts the brief only; the rep reads every brief and writes the email in the rep's own words. - Research only the companies on the rep's list, never the whole market, and count the cost per brief every week. - Clean the target list in the CRM before switch-on: one record per company, with its website and an owner. - Every line of the brief carries its link and the date it was read; a line with no link is deleted, and the count of deleted lines is an alert. - No send mode, ever: every draft waits in the rep's queue, and the machine has no email account. The map of this system, a 3x3: - The middle square, the judgment it all runs on: Which account is worth a call, and what to say (The rep, with the VP of business development). It stays with a person. - The work, in the order it moves: 1. Read the target list (runs on its own: From the CRM, every Monday); 2. Gather the facts (runs on its own: The pipeline page, the trial registry, press and filings, people by title); 3. Spot the trigger (runs on its own: Any change under 30 days old); 4. Draft the brief (runs on its own: One page, every line linked, gaps left as gaps); 5. Write the email (a person does it: The rep, in her own words, under her own name). - Along the bottom, what it runs on: Automation (The weekly schedule, the daily watch and the off switch.); Measure (Minutes a brief, lines struck, emails sent, replies, first meetings.); Build (The ideal client profile, the brief's fixed fields, which sources count.). Build the squares in that order, and keep the person's squares with a person. The working set to build from: - Trigger: Monday at 6:00, for every company on the rep's target list in the CRM; a daily check at 7:00 on the watched accounts (every company briefed in the last 30 days with no reply yet, about 100 at this volume). - Steps: Read the target list from the CRM; for each company, fetch its website's pipeline or products page; search the public trial registry by sponsor name for phase, status, start and completion dates and dosage-form words (tablet, capsule); search the last 90 days of press releases and filings for a financing, a trial start, or a manufacturing partner; find people with CMC, manufacturing, technical operations or supply chain titles, name and title only, from public pages; draft the brief in fixed fields (product and dosage form; stage; timeline; decision makers; likely manufacturing need; triggers under 30 days old), each line with its link and read date, gaps left as gaps; flag the account if a trigger is under 30 days old; attach the brief to the company record as a note; put it in the rep's queue; on the daily run, compare each watched account with its last pull and list what changed. - Data: One record per brief: company, date, each line with its link and read date, trigger flag, the lines the rep struck or edited, minutes from open to done, email sent (yes or no, by the rep), reply received and when, meeting booked. - Review step: The rep reads every brief before using it. The VP of business development checks five briefs a week against their links. - Alerts: A brief with a line that has no link (it is deleted, and the count is kept); a source that failed to fetch; a pipeline page and a registry entry that disagree; a watched account that changed; a brief not in the queue by 8:00 on Monday; and the wrong-but-green count, the lines the rep struck after reading a brief that looked fine. - If it breaks: What you see: briefs missing, a line pointing to a dead page, or the same text two weeks running. Check first: the target list in the CRM is current and each company has a website on its record; the search service key and the model key are valid; the registry search returns results for a known sponsor. Safe fix: pause the schedule; the rep researches by hand. - Baseline to count before switch-on: One timed week by hand: minutes a brief (time it, do not ask), briefs a week, emails sent, replies, first meetings booked, and facts found wrong after a meeting. - Tools by name: The workflow runner n8n (Starter, 20 euros a month billed annually); the search service Exa ($7 per 1,000 searches, $1 per 1,000 pages); the model Claude Sonnet 5.5 ($2 per million tokens in, $10 out); the public trial registry ClinicalTrials.gov and the SEC's EDGAR full-text search for filings, both free; the CRM's API (HubSpot or Salesforce, the posting prefers either). Inside the CRM: HubSpot Sales Hub Professional ($90 a seat a month annual, $1,500 onboarding, 3,000 credits a month, more at $0.010 a credit) with its prospecting agent, which spends credits per lead it writes outreach for and has a send-without-review mode this build never uses. Standalone: Apollo Professional ($79 a seat a month annual, 48,000 credits a seat a year, AI research); Clay (from $60 a month at the smallest paid tier, $185 headline; data credits and actions). People: Stealth Agents (from $1,600 a month, full time). It may never, on its own: Send or schedule an email; guess an email address, a timeline or a decision maker; write a line without a link; change a CRM field; or read or store anything a prospect sent us in confidence. Keep with a person: Which accounts go on the list, because only a person knows what the company can make; the angle; the first email, in the rep's own words under her own name; the call; which line to trust when two sources disagree; and anything a prospect has said in confidence. Not yet: Sending any email on its own; guessing an email address, a timeline or a decision maker the sources do not state; scoring an account as ready; writing to any CRM field other than attaching the brief as a note. Ask me for the tools and numbers you need from my business before you start. Then build it one step at a time, and show me each step working before the next.
No one in your business holds this judgment?
Then the first step is the expertise, not the build. Konstellation, Nick's advisory firm, judges one of your jobs with you, free: what it costs today, the verdict, the bill and who owns it.
Get the verdict on one jobKnow someone doing this job by hand? Send them this page.
Free, for one email
One email opens every job and every system.
Everything on instig8.ai is free. Your email, once, unlocks all of it on this device, and you hear when the next job is out.
Unsubscribe any time.