AI Sales Automation: What an Agency Actually Builds

Most sales teams already pay for a CRM, a sales engagement tool, an enrichment vendor, a meeting scheduler, and at least one AI writing add-on. The pipeline is still leaky, the reps still spend half their week on admin, and forecast accuracy is still a guess. An AI sales automation agency is hired to fix the gap between the tools you own and the outcomes the board expected when it bought them.

This is a working guide to what those agencies build, what they charge, what the stack typically looks like, and how to scope a first engagement so it pays back inside the financial year. The mid-market context (50-1000 employees, HubSpot or Salesforce, a handful of integrations already in place) is the default lens, but most of this transfers up or down a tier.

What an AI sales automation agency actually builds

Strip away the marketing language and the work falls into six concrete buckets. Most engagements touch three or four of them.

Lead scoring and routing. Replacing the points-based scoring inherited from the CRM template with a model trained on closed-won and closed-lost data. The output is a probability, not a band, and it feeds routing rules so high-intent leads hit a senior AE within minutes rather than sitting in a queue. Done properly, this includes a feedback loop: the model retrains on outcomes, and reps can mark scores as wrong so the next version learns.

Outbound sequencing and personalisation at scale. Not generic GPT-written cold emails. The useful pattern is an enrichment-plus-research step that pulls public signals (recent funding, hiring patterns, product launches, 10-K language for US targets, Companies House filings for UK), generates a hypothesis about why this account might buy, and writes the opener around that hypothesis. The agency owns the prompt engineering, the evaluation harness, and the guardrails against the obvious failure modes (hallucinated facts, repetitive openers, mismatched tone).

CRM hygiene and enrichment. Most CRMs are 30-60% accurate on fields that matter. Agencies build pipelines that enrich on creation, dedupe on cadence, and flag stale records. The unsexy work that compounds: every other automation downstream gets more reliable when the CRM is clean.

Meeting prep and call intelligence. Pulling Gong or Chorus transcripts into a summarisation pipeline, generating pre-call briefs from CRM history plus public signals, and writing structured follow-ups straight back into the opportunity record. This is the highest-impact area for senior AEs because it removes the prep tax on every call.

Forecasting and pipeline analysis. Models that look at deal velocity, engagement signals (email opens, meeting attendance, mutual action plan progression), and historical patterns to call deals more accurately than the AE's gut feel. Useful for RevOps, sceptical-but-grateful for sales leadership.

Internal sales assistants. A RAG-grounded chatbot trained on the company's pricing logic, product documentation, competitor battlecards, and approved messaging. Reps ask it questions instead of pinging product marketing on Slack. Saves hours per rep per week in organisations large enough to have the documentation problem.

The stack: what gets used and why

There is no single right stack, but the choices cluster. A reasonable mid-market build in 2026 looks like this:

CRM as the system of record. HubSpot for under 200 reps, Salesforce above that, with the usual exceptions for industry-specific platforms (Veeva in life sciences, nCino in banking). The CRM holds the truth; automations read from it and write back to it.

Workflow orchestration. n8n self-hosted is the default for anything that touches sensitive data or runs frequently enough that Zapier task pricing becomes painful. Zapier or Make for quick wins and lighter integrations. Custom Python or TypeScript when reliability and observability matter more than visual editing - typically for anything in the critical revenue path.

LLM layer. OpenAI's GPT-4o or Anthropic's Claude Sonnet for general reasoning and writing, with Claude preferred for longer-context work and OpenAI preferred where function-calling reliability matters. Smaller open-source models (Llama 3, Mistral) for high-volume, low-complexity tasks where cost per call dominates - classification, extraction, basic summarisation.

Retrieval for grounded assistants. Postgres with pgvector is the pragmatic default. Pinecone, Weaviate, or Qdrant when scale or specific features justify the extra vendor. Hybrid retrieval (BM25 plus dense vectors) for anything where exact terminology matters - product SKUs, contract clauses, pricing tiers.

Enrichment and signals. Clay for orchestrated enrichment, Apollo or ZoomInfo for contact data, Common Room or Koala for product-led signals, LinkedIn Sales Navigator where compliant scraping is in scope (note: ICO guidance on lawful basis still applies, especially for B2B outreach to named individuals).

Call intelligence. Gong or Chorus for established teams, Fireflies or Fathom for smaller budgets. The integration layer matters more than the vendor choice - getting transcripts out reliably is the agency's job.

What it costs and what payback looks like

Honest numbers, based on what UK mid-market agencies charge in 2026:

Initial build: £15k-£60k for a single workstream (lead scoring, or outbound automation, or an internal assistant). £60k-£200k for an integrated rollout touching three or four of the six buckets above. Anything quoted under £10k for a real custom build is either a productised template or someone underestimating their own time.

Ongoing operation: £3k-£12k per month for retained support, depending on how many automations are in production and how much iteration the business wants. Most mid-market clients continue on retainer because LLM behaviour drifts, vendor APIs change, and the business itself evolves.

LLM and infrastructure costs: Usually £500-£5,000 per month in usage depending on volume. A 30-rep team running personalised outbound and call intelligence typically lands around £1,500-£2,500 monthly in API and hosting costs.

Payback is the part to argue about before signing. A defensible target is one full-time-equivalent of recovered selling time per 10 reps within six months, plus a measurable lift in connect rate or meeting-booked rate on automated outbound (10-30% is the realistic band, not the 3x claims you see in vendor case studies). McKinsey's 2024 sales research puts the productivity uplift from generative AI in B2B sales at 10-15% of total seller capacity, which matches what we see in practice once the system is bedded in.

If an agency cannot tell you, in writing, what metric will move and by how much, that is a scoping problem, not a technology problem.

How to scope a first project that ships

The first project should be small enough to ship in 8-12 weeks and large enough to be felt by the sales team. The pattern that works:

Week 1-2: discovery. The agency interviews 4-6 reps, two managers, and the RevOps lead. They pull data from the CRM, the sales engagement tool, and the call intelligence platform if one exists. The output is a written diagnosis of where time is being lost and where the most credible early wins sit. Expect to pay for this whether or not you proceed - serious agencies do not give it away.

Week 3-4: specification. A single workstream chosen. Success metric agreed in writing. Data access confirmed. The DPIA started if any of this touches personal data of UK or EU residents (it almost always does for outbound).

Week 5-10: build and iterate. Working software in front of real reps by week 6 at the latest. Weekly demos. The agency should be running an evaluation harness against the LLM outputs - not just shipping prompts and hoping.

Week 11-12: handover and instrumentation. Dashboards for the metric you agreed to move. Runbook for what breaks and who fixes it. Training for the RevOps team if they will operate it, or a retainer agreement if the agency will.

The single biggest failure mode is scoping too much for the first build. Two or three buckets in parallel usually means none of them lands. One bucket done well builds the political capital for the next.

Agency vs in-house vs platform

The honest comparison:

In-house build. Works if you already have a RevOps team with engineering capacity, or an internal data science function with spare cycles. Cheaper over a three-year horizon if the team stays. Slower to ship the first system, and the opportunity cost of the engineers' time is often underestimated. For most mid-market companies, in-house becomes viable after the first agency-led build has proved the pattern.

Platform-only. HubSpot's AI features, Salesforce's Einstein, Apollo's AI sequencing, Gong's deal intelligence. These are real and improving fast. The limit is that they work on the data and workflows the platform already sees. Anything custom - your specific qualification logic, your odd industry signals, your integration with the legacy quoting system - sits outside what the platform can do. Most mid-market companies end up with platform AI for the obvious bits and custom builds for the differentiated bits.

Agency build. Faster to first value, more expensive in year one, and you avoid the hiring market for AI engineers (which is brutal in 2026). The risk is vendor lock-in: pick an agency that writes code you can read, hosts on infrastructure you can take over, and documents the system properly. Ask to see the runbook from a previous engagement before signing.

A pragmatic stack for most mid-market businesses: platform AI for the 60% that is commodity, agency-built systems for the 30% that is differentiated, and a small in-house RevOps capability for the 10% of daily tweaks and operation.

What to ask an agency before you hire one

The questions that separate practitioners from PowerPoint:

• Show me a system you built that is still in production 18 months later. Who operates it now?
• How do you evaluate LLM outputs before they go live? What does your evaluation harness look like?
• What is your default approach to PII and the UK GDPR? Where does data sit?
• What happens when OpenAI or Anthropic deprecates the model you built against?
• Show me the runbook from a recent handover.
• What is your retention rate on retainer clients past 12 months?
• How do you price change requests after the initial build?

If the answers are vague, the work will be too. Good agencies have rehearsed answers to all of these because their best clients asked them.

Frequently asked questions

How long does it take to see results from AI sales automation?

First measurable results within 8-12 weeks for a well-scoped single workstream. Lead scoring and CRM enrichment tend to show movement fastest because the baseline is usually so poor. Outbound personalisation takes longer to read because you need enough volume to see statistical significance in reply rates - typically 6-8 weeks of sending after launch. Internal assistants show qualitative wins immediately (reps stop asking the same questions on Slack) but the quantitative time-savings take a quarter to land in any management dashboard. Plan for a full quarter of operation before reviewing whether to expand scope.

Will an AI sales automation agency replace our SDRs?

No, and any agency that pitches it that way is selling something they will not deliver. The realistic pattern is that SDRs become responsible for more accounts each, with the automation handling the research, first-touch personalisation, and follow-up cadence while humans handle the conversations that actually need a human. Teams typically see 30-50% more accounts covered per SDR, not a headcount reduction. The companies that have tried full replacement have mostly walked it back within a year because reply quality collapses and brand damage is hard to reverse.

What about UK GDPR and B2B outreach?

UK GDPR applies to personal data of UK residents even in B2B contexts, particularly when you are processing named individuals' email addresses, job titles, and inferred interests. The lawful basis for B2B outbound is usually legitimate interests, which requires a documented legitimate interests assessment (LIA). Any agency building outbound automation should be doing a DPIA as part of scoping, configuring retention policies in the enrichment tooling, and respecting opt-outs across systems. The ICO's direct marketing guidance is the reference document. PECR also applies for electronic marketing and has separate rules.

Can we self-host the AI components for data sensitivity reasons?

Partially yes. n8n, Postgres with pgvector, and open-source models (Llama 3, Mistral) can all run in your own cloud account or on-premise. The LLM layer is where it gets harder: GPT-4-class quality from a self-hosted open model is closing the gap but still 6-12 months behind the frontier in our experience. For most sales automation use cases, the pragmatic answer is self-host the orchestration and data layer, use OpenAI or Anthropic via their enterprise tier with data processing agreements in place, and accept that the frontier models are not going on your own GPUs in 2026 unless you have a very specific reason.

How do we avoid vendor lock-in with an agency?

Three things to insist on in the contract. First, code and infrastructure-as-code lives in your repos and your cloud account from day one, not the agency's. Second, the agency documents the system as part of the deliverable - runbooks, architecture diagrams, prompt libraries with version history. Third, you have the right to engage a different agency or in-house team to operate the system from any renewal point with a reasonable handover period (30-60 days is normal). Agencies that resist any of these are protecting recurring revenue at your expense. The good ones offer it before you ask.

What size sales team makes this worth doing?

The honest floor is around 5-8 reps. Below that, the build cost rarely amortises in a sensible timeframe and platform AI features cover most of what you need. Between 8 and 30 reps is the sweet spot - large enough that automation compounds, small enough that you can ship changes without months of internal politics. Above 30 reps the prize is larger but the implementation gets more complex because you are touching more processes and more stakeholders. Enterprise sales teams (100+ reps) usually need a phased rollout by segment or region rather than a single project.

What if the LLM gets something wrong and sends a bad email?

This is the right question to ask, and the answer is design for it rather than pretend it cannot happen. Production-grade systems have three layers of defence: an evaluation harness that scores outputs against rubrics before they go live, a human-in-the-loop step for anything novel or high-stakes (typically the first email to enterprise accounts above a deal-size threshold), and monitoring that flags outputs which look statistically unusual. A small percentage of edge cases will still slip through. Reps should be trained to spot and escalate them, and you should have a process for incident review that feeds back into the prompts and guardrails.

How does this work with our existing HubSpot or Salesforce setup?

The CRM remains the system of record. Automations read from it, enrich and process, and write back to it. The agency should be working in your sandbox first, with field-level mapping documented before anything touches production. Custom fields and objects are added carefully and reversibly. Workflows the CRM already runs are mapped and either kept, replaced, or augmented - never silently overridden. Expect a week of integration work upfront just to get the read-write patterns right and the field naming aligned with whatever conventions your RevOps team has built up over the years.

Where to go from here

The mid-market companies getting real value from AI in sales are the ones who picked one workstream, shipped it properly, measured it honestly, and then expanded. The ones still on slide 47 of an AI strategy deck two years in are the ones who tried to do everything at once or hired a strategy firm that does not build. If you want to talk through where the highest-impact opening sits in your sales operation, get in touch with AI Advisory - we run a fixed-fee two-week readiness engagement that ends with a costed roadmap and a recommendation on what to build first.