A8gent
Build Proof Assets Clients Trust cover
Offer Design · Lesson 2 of 15

Build Proof Assets Clients Trust

Clients buy proof, not claims. Learn which assets show an agent workflow works before a prospect has paid you anything, and how to build each one fast.

Proof closes deals that claims cannot

A marketing agency selling AI agents faces a trust problem: the prospect has heard the hype and seen tools that underdelivered. You close by showing proof that this specific workflow produces a real result, not by describing capabilities. Proof assets are the concrete artifacts a prospect can look at and believe. Every hour spent building them pays back across every future sales call, because the same pack sells the same offer again and again.

The proof assets that work

  • A worked before-and-after. Take one real task, show the manual process and time, then show the agent output next to it. A weekly report that took an analyst two hours drafted in minutes is more persuasive than any slide.
  • A recorded walkthrough. A short screen recording of the workflow running on real-looking data, including the human review step, shows the client where they stay in control.
  • A sample output pack. Redacted example reports, drafted replies, or QA flags the prospect can hold in their hands and judge against their own standards.
  • A clear scope one-pager. What the agent does, what it does not do, and where a human approves. This reassures serious buyers as much as the output itself.

The before-and-after evidence table

The strongest single page in your proof pack is a measured comparison. Build it from a real run of the workflow, not estimates, and put it at the front of every proposal:

Metric Manual process (measured) With agent (measured)
Time per weekly report2h 10m analyst time12m review time
Reports delivered on time14 of 20 by Monday noon20 of 20 by Monday 9am
Data errors caught in reviewUnknown, found by clientsFlagged before send, logged
Who approves before sendingAccount managerAccount manager (unchanged)

Note the last row. Showing that the approval step is unchanged is itself proof: it tells a cautious buyer that speed did not come at the cost of control.

Build proof from your own agency first

You do not need a paying client to create proof. Run the workflow on your own agency's reporting, lead follow-up, or content production. This gives you honest numbers, a real recording, and the experience to answer hard questions. It also means you believe your own offer before you sell it. If your own team will not use the workflow weekly, that is a signal to fix before selling it to anyone.

Keep the pack honest and current

A proof pack is only as good as its worst claim. Re-run the numbers every quarter and after every meaningful workflow change, and date every figure so a prospect can see the evidence is fresh. When recording the walkthrough, resist the urge to cut the slow parts: a two-minute clip that includes the human clicking approve is more convincing than a slick thirty-second montage, because the approval click is the part a cautious buyer is looking for. And when a real client gives you permission to use their numbers, swap out your internal example immediately; third-party proof outranks self-generated proof every time.

The scope one-pager template

Serious buyers ask "what happens when it gets something wrong" before they ask about price. Answer it in writing before they ask:

SCOPE ONE-PAGER - Weekly Reporting Agent

WHAT IT DOES
- Reads: Google Ads, GA4, Meta Ads for connected accounts
- Produces: one drafted weekly report per account, in your template
- Runs: every Monday 6am, drafts ready by 9am

WHAT A HUMAN STILL DOES
- Account manager reviews every draft before it is sent
- Any anomaly the agent flags is decided by a person
- Nothing reaches your clients without an approval click

WHAT IT DOES NOT DO
- No strategy recommendations
- No changes to campaigns, budgets, or bids
- No direct sending to your clients

WHEN IT IS UNSURE
- Missing or conflicting data is flagged, never invented
- Flagged reports are held for human review with a reason attached

Which proof asset for which buyer

Buyer situation Lead with Why
Skeptical, burned by AI tools beforeSample output packLets them judge real work against their own standards
Interested but time-poor ownerBefore-and-after tableOne page of measured numbers, decision in minutes
Cautious about brand and controlScope one-pager + walkthroughShows the approval step and the hard limits first
Technical evaluator on the callRecorded walkthroughShows the real workflow, not slides

How to present proof in a call

  1. Open with the client's current painful process, in their words.
  2. Show the before-and-after on a task that looks like theirs.
  3. Point to the human review step so they see the agent does not act unchecked.
  4. Hand over the sample output and let them critique it, then note that reviewer edits are exactly how the workflow improves.

What good looks like

A prospect can look at your proof and picture the workflow running in their own agency by Friday. They are reacting to a real output, not a promise. That shift, from claim to artifact, is what separates agencies that sell agent work from those that only talk about it.

Your first proof, when you have delivered nothing

The uncomfortable position everyone starts in: no clients for this offer, no case study, and a prospect asking who else you have done this for. There are four honest routes out of it and one dishonest one. Take any of the four.

  1. Run it on your own agency, this week. The single best first move, and the one most people skip because it feels like it does not count. It does. You get real timings, a real recording, and - most valuably - the experience of the thing going wrong once, which is what lets you answer hard questions calmly. "We run this on our own reporting every Monday" is a strong sentence and it is true.
  2. Build it once at a reduced price, deliberately. Not free, and not a favour - a stated trade. "This is the first build of this offer. It is $1,500 instead of $4,500. In exchange I need permission to use the numbers and a fifteen-minute conversation on camera when it is working." Put the exchange in writing so nobody is confused later about what was bought.
  3. Build a public sample on public data. Ad platform data is private, but plenty of workflows can be demonstrated on something anyone can see. A QA agent can run over a public site's UTMs. A reporting agent can run over a demo analytics account. Nobody mistakes it for a client result, and it proves the mechanism.
  4. Borrow proof from the process, not the outcome. Your scope one-pager and your discovery question set are themselves evidence - they show a buyer that you have thought about failure modes, approval, and limits. Cautious buyers read that as competence, which is exactly the thing they are trying to assess.

The dishonest route is inventing a case study, an anonymous client, or a percentage. It is also the one that gets discovered: a buyer asks one follow-up question about a client you do not have, and the entire conversation is over along with your reputation in a small vertical. The four routes above take a fortnight and cost nothing you cannot afford.

Measuring the "before" properly

The before-and-after table is only proof if the before is real, and the before is the half people fabricate without meaning to - by estimating, from memory, in their own favour. Measure it like this instead:

MEASURING THE MANUAL BASELINE

Do this BEFORE you build anything. It is unrecoverable afterwards.

1. Pick the next 5 real runs of the task.
2. For each one record, at the time and not from memory:
     - who did it
     - start and end timestamp (including interruptions)
     - anything that went wrong or had to be redone
     - whether it was delivered on time
3. Report the MEDIAN, not the average, and say so.
   One disastrous run should not flatter your after-figure.
4. Write down what you did NOT measure. For example:
     - the account manager's review time
     - the time spent chasing missing data
   Then either measure those too, or state plainly that
   they are excluded from both columns.

Step four is the one that makes a table credible to a sceptical buyer. A comparison that quietly excludes review time from the "after" column and includes it in the "before" column is the most common way an honest person produces a dishonest number. Say what is excluded, exclude it from both sides, and the table survives scrutiny - which matters, because the buyers most worth having are the ones who scrutinise.

Turning a delivered build into a case study

The window for this is narrow: two weeks after go-live, while the client still remembers the before. Wait three months and you get "yeah, it's been good", which is worth nothing.

Ask these five questions, record the answers, and get written permission before you use anything:

  • "What was the Monday morning like before this?" Their description of the pain, in their words, is the opening line of the case study.
  • "What is it like now?" Resist paraphrasing. The client's own understated phrasing is more credible than anything you would write.
  • "What did you expect to go wrong that didn't?" This surfaces the objection your next prospect has and has not said out loud.
  • "What did go wrong?" Include it. A case study with one honest wrinkle and how it was handled outperforms a flawless one, because flawless reads as marketing.
  • "Would you say that on a fifteen-minute call to someone considering this?" A referenceable client beats a written quote by a distance.

On permission: get it in writing, be specific about what you may use - name, logo, numbers, quote - and accept a redacted version gladly. "A paid media agency managing around 30 accounts" with real numbers is far stronger than a named logo with vague ones.

Building the sample output pack, item by item

"Sample output pack" is the asset people describe vaguely and then assemble badly - usually three cherry-picked perfect outputs, which a sceptical buyer discounts on sight. A pack that persuades has a specific composition, and it is deliberately not all good news.

SAMPLE OUTPUT PACK - Weekly Reporting Agent
Six files. Assembled once, reused on every call.

01-typical.pdf          A completely ordinary week. Boring on
                        purpose. This is the one they will judge.

02-messy-week.pdf       A week where a data source was late.
                        Shows the flag, the held report, and the
                        note explaining what is missing.

03-flagged-anomaly.pdf  A week where spend jumped. Shows the agent
                        surfacing it rather than explaining it
                        away - and NOT recommending an action.

04-before-after.pdf     The measured comparison table, dated, with
                        the exclusions stated at the bottom.

05-reviewer-diff.pdf    One report as drafted, and the same report
                        as sent, with the account manager's edits
                        highlighted. See below - this is the one
                        that closes cautious buyers.

06-scope.pdf            The scope one-pager. What it does, what a
                        human does, what it never does.

File 05 is the counterintuitive one, and it is the strongest item in the pack. Every instinct says to hide the edits, because edits look like the agent failed. They do the opposite. A buyer who sees a draft with four small corrections marked on it learns three things at once: the review step is genuinely happening, a competent human is exercising judgment over the output, and the volume of correction is small enough to be worth it. A pack of flawless outputs raises the question "who checked these?" and leaves it unanswered.

On redaction, the rule that keeps you out of trouble: redact identity, never redact structure. Replace the client name, account IDs and absolute spend figures; keep the layout, the metric names, the flags, the annotations, and the relative movements. A buyer needs to see the shape of the work, not your other client's budget. Do the redaction on a copy, in a way that cannot be undone - retyping into a clean template, not drawing black boxes over a PDF, which frequently leaves the text underneath selectable.

The walkthrough recording, as a shot list

Most recorded walkthroughs fail the same way: they are a tour of a tool rather than a demonstration of a job being done. Record against this list instead. Under four minutes, one take, no editing beyond trimming the ends.

Segment What is on screen What you say, roughly
0:00-0:20The manual process. A spreadsheet, half-built, tabs open"This is what Monday looks like now. Three platforms, one deck, about two hours."
0:20-0:50The trigger firing, unedited"Nobody starts this. It runs at six."
0:50-1:40The draft appearing. Let the wait be visibleSay nothing for a few seconds. Real duration reads as real
1:40-2:40The reviewer opening it and changing two things"Sam checks this. Here she's fixing a label and cutting a line."
2:40-3:10The approve click, and where it goes"Nothing moves until that click. That doesn't change."
3:10-3:40A held report with a missing-data flag"When something's missing it stops and says so. It doesn't fill the gap."

Two production notes that matter more than the camera quality. Do not cut the waiting. A montage that compresses a ninety-second run into four seconds reads as a sales video; the unedited wait reads as a real system. And end on the failure segment, not the success segment. The last thing a cautious buyer sees should be the agent declining to guess, because that is the fear they came in with, and answering it last is what they remember.

Troubleshooting: proof that is not landing

What happens on the call What it actually means What to change
"Impressive, but our setup is different"The sample does not look like their workRebuild the pack inside the vertical you chose. Generic proof persuades nobody in particular
They go quiet at the numbersThe comparison looks too good, so they assume it is riggedAdd the exclusions line and the median note. Credibility rises as the claim shrinks
"Who checks it?" asked more than onceThe review step was mentioned but never shownLead with the reviewer-diff file, before the outputs
They ask for a trial on their own dataA buying signal, and a scope riskSay yes as a paid, time-boxed first build with a written finish line, not as a free pilot with no end
Enthusiasm on the call, silence afterwardsThey cannot re-explain the offer to whoever holds the budgetLeave the one-line pitch and the before-after page as a two-page leave-behind. Your buyer has to sell it without you
They fixate on a single flaw in a sampleNormal, and usefulAgree, and say that reviewer edits are how the rules get written. Defending the flaw is the only losing move

The fifth row is the one that quietly loses the most deals. The person you present to is very often not the person who approves the spend, and your proof pack has to survive being described second-hand by someone less fluent than you. That is the entire argument for the one-line pitch and the single-page comparison: they are not simplifications for the buyer, they are ammunition for the buyer.

When proof assets are not the bottleneck

  • You have no offer yet. Proof of a vague offer is just a demo. Fix the offer definition first - a pack built around a moving target has to be rebuilt every time it moves.
  • The prospect already trusts you. An existing retainer client who likes your work does not need a pack; they need a scope one-pager and a price. Do not perform a sales process at someone who is already sold.
  • The blocker is procurement, not belief. If the hold-up is security review or data processing terms, more samples change nothing. Get the scope one-pager, your data handling in writing, and whatever their supplier form asks for - that is the actual bottleneck.
  • You cannot deliver it yet. Proof that outruns your delivery capability is a fast route to a client who bought something you cannot build. Build the thing, then prove it.

A two-week schedule for building the whole pack

The pack is six files, a recording and a page of numbers, and it will sit undone for a quarter unless it is scheduled. Here is the version that fits around delivery work, assuming you are running the workflow on your own agency first.

When What you do Time
Day 1Start the baseline log. Five real manual runs, timed as they happen. Nothing else - this is the only step that expires10 min setup
Days 1-10Baseline log fills in the background while you buildpassive
Day 8Write the scope one-pager. Do it before the outputs exist - it forces the limits to be decisions rather than descriptions of whatever you happened to build45 min
Day 11Collect the six sample files from real runs. Do not generate them specially - use what actually came out, including the messy week1 hour
Day 12Redact by retyping into a clean template. Never by drawing boxes over a PDF1 hour
Day 13Record the walkthrough against the shot list. One take. Re-record only if something is genuinely inaudible30 min
Day 14Build the before-and-after page from the baseline log. Median, dated, exclusions stated30 min

Under five hours of actual work, spread over a fortnight, and the ordering matters more than the total. Day 1 is first because a manual baseline cannot be reconstructed once the agent exists - every day you delay it, the memory of the old process gets vaguer and more flattering. Day 8 is before day 11 because a scope one-pager written after the build is a description of what you happened to make, whereas one written before it is a commitment you then build to. Everything else can slip.

One scheduling note decides whether this happens at all: put day 1 and day 8 in the calendar as appointments with yourself, and leave the rest as a checklist. Those two are the only steps that cannot be recovered later - the baseline expires the moment the agent exists, and a scope one-pager written after the build is a description rather than a commitment. Everything from day 11 onward is assembly work that will wait a week without losing value, so protecting only the two irreversible steps is enough to get the whole pack built. Founders who try to protect all seven days protect none of them.

The other reason to start the baseline before the build is that it changes what you build. Timing five real runs forces you to watch the manual process closely, and you will invariably find that a step you assumed was the expensive one is not - the two hours are going on chasing a missing data source, or on reformatting, rather than on the analysis you were about to automate. That discovery is worth more than the numbers themselves, and it only happens if the measuring comes first.

Common mistakes

  • Using invented numbers or polished demo data that collapses the moment a real client tests it.
  • Showing capability slides instead of a single concrete output the buyer can judge.
  • Hiding the human review step, which makes cautious clients trust the whole thing less.
  • Building a new proof pack per prospect instead of one reusable pack per offer.

What you should be able to do now

  • Name the four proof assets and what each one convinces a buyer of.
  • Build a measured before-and-after table from a real run of your workflow.
  • Write a scope one-pager that answers "what happens when it is wrong" in advance.
  • Match the right proof asset to a skeptical, time-poor, cautious, or technical buyer.
  • Run the four-step proof presentation in a sales call without capability slides.
  • Measure a manual baseline honestly, including stating what you excluded from both columns.

Where this goes next

This is the second free lesson. With the offer lesson you now have a narrow offer and an honest way to prove it works before anyone has paid you. The single highest-value thing you can do this week is measure the manual baseline on your own agency's reporting - it takes five runs and it is unrecoverable once you have built anything.

The paid lessons build the thing the proof pack describes. The reporting agent end to end, including the data joins that break in practice. Brand-voice systems that hold up across more than one writer. Review workflows so nothing reaches a client unchecked. Pricing for agencies, selling into an existing retainer, and the proposal that closes. Then delivery at volume - templates, SOPs, subcontracting, and QA that keeps quality flat as the client count rises. Every lesson ships the workflow files and the documents.

If your read is that you want the system rather than the skill set, we build these as projects - see AI automation services, or take the free automation audit to work out which delivery workflow is worth automating first. For the commercial background, AI automation cost and the proposal template are both full documents rather than teasers.

Put your email in below and the free lessons arrive with the scope one-pager, the baseline measurement sheet, and the case study question set as editable files - the whole proof pack skeleton, ready to fill in.

Key takeaways

What to remember from this lesson

  • Four proof assets do different jobs: a measured before-and-after, a recorded walkthrough, a sample output pack, and a scope one-pager.
  • Build proof from your own agency first. If your own team will not use the workflow weekly, fix that before selling it to anyone.
  • Show the approval step, do not hide it. A clip that includes the human clicking approve beats a slick montage, because that click is what a cautious buyer is looking for.
  • Answer 'what happens when it gets something wrong' in writing before they ask - serious buyers ask it before they ask about price.
  • Date every figure, re-run the numbers quarterly, and swap in third-party proof the moment a client permits it. Third-party proof outranks self-generated proof every time.

Try it yourself

Build the before-and-after row from a real baseline log

Here is a baseline log: five timed manual runs of a weekly report, and five timed review sessions once the agent was live. These are made-up timings, not measured results from anyone's agency. Build the comparison row honestly. The lesson says report the median and say so, and this fixture contains the two traps that catch people: one disastrous run, and a sort that goes wrong.

Write it

Write the reduced-price first build as a stated trade

Route two in the lesson is building it once at a reduced price - not free, and not a favour, but a written exchange. Write the message you send. This checks structure only: whether the trade is priced, bounded, and states what you get back. It cannot tell you whether the prospect will say yes.

2 of 7 checks passing

  • Says plainly that this is the first build of the offer
  • Names both numbers, the reduced price and the standard one
  • States what you get in exchange
  • Asks for permission to use the numbers or a reference conversation
  • Puts it in writing rather than leaving it a favour
  • Never sells it as free or as a favour
  • Short enough to be read in one go

These checks look at structure only, not quality. A prompt can pass every one and still be a poor prompt, so read it back and ask whether you would understand it cold.

Check yourself

Which asset, and why the counterintuitive one works

  1. A prospect who was burned by an AI tool last year joins the call. Which asset do you lead with?

  2. Why does the reviewer-diff file - a draft shown next to the edited version that was sent - close cautious buyers?

  3. You measured the manual process at two hours and the agent version at twelve minutes, but the twelve minutes excludes the account manager's review while the two hours includes it. What is the fix?

Answer every question to mark.

Exercise

Build one reusable proof pack from a real run

120 min
  1. 1Pick one workflow and run it on your own agency's work, not a client's, so you can publish the numbers without permission.
  2. 2Before automating it, time the manual version at least twice and record the actual minutes. Estimates do not survive a prospect's questions.
  3. 3Run the agent version and record the review time. Build the before-and-after table with at least four rows, and date it.
  4. 4Include the row that shows who approves before sending, with the same answer in both columns. That row is proof that speed did not cost control.
  5. 5Record a walkthrough on real-looking data and keep the human review step in the recording, uncut.
  6. 6Write the scope one-pager with all four blocks: what it does, what a human still does, what it does not do, and what happens when it is unsure.
  7. 7Assemble two or three redacted sample outputs a prospect can critique against their own standards.
  8. 8Match each asset to a buyer type using the table, so you know which one to lead with on a call.

You are done when Your pack contains four dated assets - a before-and-after table whose numbers came from a timed run rather than an estimate, an uncut walkthrough that shows the approval click, a scope one-pager with all four blocks, and redacted sample outputs - and it is built once for the offer, not per prospect.

Companion files

Scope one-pager and evidence table templates

The scope one-pager structure, the measured before-and-after table, and the buyer-to-asset matching table - the fill-in versions of the artifacts this exercise asks you to produce.

The free-preview pack includes the worksheet for this lesson for Build Proof Assets Clients Trust, plus the other free lessons of this course. No email needed.

Free preview pack - 4 KB zip

Check yourself

Answer before you reveal

You do not have a paying client yet. Can you build a proof pack?Show answer

Yes, and you should. Run the workflow on your own agency's reporting, lead follow-up, or content production. It gives you honest numbers, a real recording, and the experience to answer hard questions - and it tests whether you believe your own offer.

A skeptical prospect who was burned by an AI tool joins the call. Which asset do you lead with, and why not the before-and-after table?Show answer

The sample output pack. A burned buyer trusts their own judgement, not your numbers - hand them real work and let them critique it against their standards. Numbers land better with a time-poor owner who wants to decide in minutes.

Why does hiding the human review step backfire?Show answer

It makes cautious clients trust the whole thing less. Their fear is unchecked action; showing the approval gate is the thing that removes it. Concealing it reads as either a missing control or something you are not saying.

Free lesson notes

Get the notes pack for AI Agents for Marketing Agencies Course, plus every new module free by email.

No card, no spam, unsubscribe any time.