AI automation audit · sample
Northwind Ops
- company
- 65 ppl, €9M ARR
- date
- May 2026
- prepared by
- amberstack
executive summary
We ran the audit with Olga K. (COO, signer), 30-min kickoff on day 2, then 60-min interviews with three teams: Sales (11), Customer Success/Support (12), Engineering (22). Total 19 hours of amberstack work across 10 working days.
Five candidate workflows surfaced. One lands in Phase 1 (≤90 days payback) at the €11,000 single-workflow build price: outbound prospect research + personalization briefs. Saves ~273 h/month of SDR time (240h outbound prep + 33h inbound enrichment side-effect), net €6,432/month, payback ~52 days.
Two more land in Phase 2 (91–180 days payback) - recommended after Phase 1 ships and trust is established. Two are flagged Hold (>180 days on hours-saved math alone). The math is in §4.
Three items came out of interviews as “don't automate” - most importantly, the outbound email itself (Phase 1 stops at the research brief). Reasoning in §5.
Recommended next step: fixed-bid Phase-1 build for €11,000, kickoff the week of June 9, Operations retainer at €2,200/month following go-live (90-day initial term, monthly thereafter). SOW available within 48 hours of this readout.
1 · context
1.1 Business goal this audit serves
Northwind is targeting €14M ARR by Q4 2026 without growing headcount past 75. The board narrative is “capital-efficient growth from outbound.” The COO identified Sales prospecting capacity as the binding constraint - current SDR throughput limits how many ICP accounts get touched per quarter.
1.2 What “good” looks like
- Sales: each SDR handles 2× current outbound touches without dropping personalization quality.
- Time-to-first-touch on inbound stays under 10 minutes.
- Zero added on-call burden on Engineering from any automation we recommend.
1.3 What's out of scope
- No AI-generated outbound emails sent to prospects without human authorship.
- No changes to the HubSpot or Outreach stack - both standardized on.
- No work touching cardholder data (PCI scope kept narrow).
2 · method
Over two weeks: async stack inventory + pain hypotheses (questionnaire, completed May 8); 30-min kickoff with the COO; 60-min interviews with three teams; synthesis into a single ranked backlog scored on time saved, tractability, data risk, team buy-in, and stack readiness; per-workflow payback math against benchmark Baltic-SaaS loaded costs (Appendix B).
This audit is not a procurement evaluation, a security review, or a guarantee. Payback periods assume the workflow is built well and adopted by the team. Both are amberstack's job in the build phase, but the audit alone doesn't deliver them.
3 · stack and team summary
3.1 Stack inventory
| CRM | HubSpot |
| Sales engagement | Apollo + Outreach |
| Customer support | Intercom |
| Product analytics | Mixpanel |
| Data warehouse | BigQuery (lightly used) |
| Docs / knowledge | Notion |
| Engineering | GitHub + Linear |
| Billing | Stripe + Chargebee |
3.2 Teams interviewed
| team | lead | size |
|---|---|---|
| Customer Success / Support | Lina M. | 12 |
| Sales (SDR + AE) | Karolis V. | 11 |
| Engineering | Marek S. | 22 |
3.3 Existing AI usage. ChatGPT Team licenses across CS and Sales (~€420/mo). Cursor on the engineering team. No production AI in any customer-facing system. One abandoned internal experiment with Intercom Fin - turned off in February over response quality.
4 · opportunity backlog
Each candidate scored 1–5 on time saved, tractability, data risk, team buy-in, stack readiness. Phase column tiers against the €11,000 single-workflow build cost: P1 ≤90 days, P2 ≤180 days, Hold >180 days.
| # | workflow | team | hrs/mo | tract | risk | buy-in | stack | payback | phase |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Outbound prospect research + personalization briefs | Sales (SDR) | 273 | 5 | 2 | 4 | 5 | ~52 days | P1 |
| 2 | Sales call notes + CRM update + follow-up draft | Sales (AE) | 86 | 4 | 2 | 4 | 4 | ~166 days | P2 |
| 3 | PR review summaries + release notes drafting | Engineering | 46 | 4 | 2 | 4 | 5 | ~148 days | P2 |
| 4 | Customer health scoring + churn alerts | Customer Success | 55 | 3 | 3 | 4 | 4 | ~265 days | Hold |
| 5 | Tier-1 support draft replies | CS / Support | 60 | 4 | 3 | 3 | 5 | ~363 days | Hold |
Note: candidates #2–#3 land in Phase 2 (91–180 days). #4–#5 are Hold - they fail the 180-day bar on hours-saved math alone, but each has a secondary justification (churn save, CSAT, support SLA) that may make them worth doing later if Phase 1 lands well.
4.3 · phase-1 recommendation
Outbound prospect research + personalization briefs
One workflow, two surfaces. Drives outbound SDR throughput; same enrichment pipeline also fires on inbound form fills, so inbound triage benefits as a side effect at no extra build cost.
current state - what happens today
- SDR pulls a prospect from the ICP list in Apollo or HubSpot.
- Opens LinkedIn, company site, recent news; reads manually.
- Checks tech-stack signals (BuiltWith / Wappalyzer / HG Insights).
- Writes a 3-bullet brief in HubSpot: relevance, hook angle, urgency signal.
- Crafts the personalized outbound message based on the brief.
Per prospect: ~15 min (12 min research + brief, 3 min message). 5 SDRs × 60 prospects/wk × 4.3 wks = 1,290 prospects/mo, rounded down to 1,200/mo for the savings math (conservative). 1,200 × 12 min research = 240 h/mo outbound prep. Inbound triage adds ~33 h/mo on enrichment.
proposed state - what changes
- Daily n8n pull from Apollo: prospects matching ICP filters; HubSpot webhook for inbound.
- Enrichment pipeline runs: firmographics, hiring signals, funding, tech stack.
- Claude generates a structured brief in JSON: why-now, hook angles, recent signal, sources.
- Brief lands in a HubSpot custom property + Slack #sales-prospecting feed card.
- SDR validates (~3 min), writes the outbound message themselves, sends via Outreach.
Per prospect: ~3 min validation. Saved ~12 min/prospect. SDR keeps authorship of the actual email (see §5.1).
architecture sketch
[Apollo: ICP prospects, daily] [HubSpot: new inbound form]
│ │ webhook
└──────────┬───────────────────┘
▼
[n8n: enrichment pipeline]
│ LinkedIn, HG Insights, news API,
│ tech-stack signals, funding events
▼
[n8n: Claude brief generator]
│ strict JSON schema:
│ { why_now, hook_angles[], recent_signal, sources[] }
▼
┌─────────────────────────────────────────┐
▼ ▼ ▼
[HubSpot: [Slack #sales- [Postgres:
custom property] prospecting feed] trace + token cost]
│
▼
SDR validates ~3 min,
writes outbound, sends via OutreachROI math
| Outbound prospects researched | 1,200 / month |
| Time saved on outbound (12 min × 1,200) | 240 h/mo |
| Time saved on inbound (enrichment side-effect) | 33 h/mo |
| Total hours saved per month | 273 h/mo |
| Loaded hourly cost - mid SDR | €24 |
| Gross monthly savings | €6,552 |
| Monthly run cost (LLM + Apollo overage + hosting) | €120 |
| Net monthly savings | €6,432 |
| Build cost (fixed) | €11,000 |
| Payback period | ~52 days |
Conservative: lower bound of SDR self-reported research time, mid-SDR rate (not blended up with AE re-work). Run cost includes Claude Haiku ~1,500 calls/mo + Apollo overage allowance + €15/mo Hetzner. €24/h is the industry-standard Baltic mid-SDR loaded cost used as a default in this sample; a real audit uses your team's actual loaded rates.
risks & mitigations
LLM fabricates company facts in the brief
Strict JSON schema with mandatory
sources[]field - every claim cites a URL or Apollo field. SDR sees sources before validating. Flagged for review if no sources.Apollo enrichment quality drops, dirties HubSpot
Enrichment writes to dedicated custom property, never overwrites existing fields. Monthly hit-rate dashboard with auto-alert on drop below 80%.
VP Sales pushes back on AI in the prospecting flow
2-week shadow mode (briefs visible but SDRs don't act). Then opt-in per SDR. Kill-switch in the Slack channel pinned message. Weekly rollup of brief quality to VP for 90 days.
5 · what we recommend NOT automating
The audit's job is also to tell you what not to build. Three items, each from the interviews or from amberstack's judgment as the auditor.
5.1. Writing the outbound message itself
why not. Northwind's voice in vertical-logistics outbound is sharper than any AI-generated line. A bad email to a 200-employee logistics CIO costs more than the SDR hour it saves. (This is why Phase 1 stops at the research brief, not the email.)
what to do instead. Keep humans writing the first touch. Use AI for the research brief, hook angles, and follow-up scheduling. SDR adds the actual sentences.
5.2. Customer-facing replies on billing disputes
why not. Surfaced in CS interviews. Wrong tone or a wrong goodwill credit damages a high-LTV relationship. The downside is large, the time saved is moderate.
what to do instead. Draft-and-review tool inside Intercom - AI proposes the response and the financial gesture, the agent edits and sends. Saves prep time, keeps human accountability.
5.3. Routing escalations from L1 to on-call engineering
why not. Eng on-call already uses a documented runbook with sub-2% false-positive rate. Inserting an AI classifier here adds latency for no measurable gain and a new failure mode.
what to do instead. Leave the existing routing in place. Instrument it so we'd know if performance changes.
6 · constraints to resolve before build
| constraint | owner (client side) | needed by |
|---|---|---|
| Apollo API key + permission to write enrichment data to HubSpot custom property | RevOps lead | Before kickoff |
| Sign-off from VP Sales (named in kickoff as the likely-skeptic) on brief schema | COO | Before kickoff |
| 2-week shadow-mode window where SDRs see briefs but don't act on them - agreement in advance | VP Sales | Week 1 of build |
7 · recommended next step
amberstack recommends starting the Phase-1 build for outbound prospect research + personalization briefs at a fixed cost of €11,000, with kickoff the week of June 9, 2026.
Following go-live, amberstack proposes an Operations retainer at €2,200/month on a 90-day initial term, auto-renewing monthly. Covers 2 workflows under maintenance, ≤30h of new build per quarter, biweekly office hour, next-business-day SLA on non-critical issues and 4h on critical.
The first 30 days of the build include a fit-check exit - if you decide it isn't working, you pay the deposit only and the build wraps with whatever's been produced. The acceptance gate at go-live is a production-readiness checklist signed off before kickoff.
Full SOW within 48 hours of this readout. Signature by June 3 keeps the kickoff window.