The five workflows we automate first in almost every SMB — and the three we deliberately leave alone.

After 184 engagements, the shape of "what to automate first" has converged. There are five workflows that turn up in almost every SMB we work with, that pay for themselves quickly, and that build the operational muscle a company needs before tackling anything more ambitious. There are three other workflows that look like obvious candidates but that we now actively decline to automate, because the cost of getting them wrong is higher than the cost of the manual labour they consume.

This piece is the short version. There are no tables of percentages or eight-tab spreadsheets. It is a working list, in the order we'd recommend tackling them, with the reasons each one earns its place — and a working list of three things we'd argue you should leave the humans on, even if a vendor pitches you the opposite.

01 · The fiveWhat we automate first, in order

01 — Lead intake + automated first response

This is the first thing we automate at almost every engagement, and it is consistently the highest-ROI single intervention. The reason is the conversion cliff that we wrote about in the benchmarks piece: if your median first-response time is more than two hours, you are losing well over half the deals you would otherwise close. Closing that gap with an automated first response — acknowledgement, two or three qualifying questions, a calendar link, and an honest expectation of when a human will follow up — is the single fastest way to move conversion rate without changing anything else in the funnel.

The reason we put this first, even ahead of obvious labour-saving plays, is that revenue improvements are visible quickly to people who weren't initially in favour of automation. The skeptical sales director who wasn't sure about "this AI stuff" sees the conversion rate climb in week three and becomes the project's internal advocate. There is no faster route to organisational buy-in than a number on the dashboard moving the right way.

02 — Inbound triage and routing

The second thing we automate is everything that arrives in a shared inbox or a "general enquiries" channel and has to be sorted. Sales versus support versus partnership versus billing, urgent versus standard, English versus Spanish, new prospect versus existing client. Most SMBs do this manually, badly, slowly, and with a lot of "wrong person initially replied, please disregard" follow-ups.

We rarely automate the response at this stage — that often lives behind a human review for trust reasons. We automate the categorisation, the priority assignment, and the routing to the correct person with all the context they need to respond competently. This eliminates the most expensive failure mode of inbound: the lead or ticket that bounces between three people across two days before the right person sees it.

03 — Document drafting from structured input

Once steps 1 and 2 are working, you have something most SMBs don't: clean, structured input about every inbound interaction. That input is what turns document drafting into a tractable automation problem. The reason vendors who skip steps 1 and 2 produce bad drafts is that they're trying to draft from messy free-text inputs; with clean structure, the drafting is well-bounded and reliable.

The candidates here are quotes, proposals, statements of work, appointment confirmations, follow-up emails, and a long tail of routine correspondence. We do not yet, at this stage, automate documents that require negotiation discretion or pricing judgment — those are explicitly on the "leave it alone" list below. We automate the routine end of the spectrum, where the document is mostly assembling known facts in a standard format.

04 — Meeting summary and CRM hygiene

This is the workflow that everyone agrees they should be doing well and that almost nobody is. Sales and support people hate writing up calls. CRMs are full of one-line entries like "good chat, will follow up", which are useless three months later. The result is a pipeline view that lies to the company about its own state.

Automating meeting summary — call transcript → structured summary → relevant CRM fields populated → suggested follow-ups created — is a workflow that takes a few weeks to build and that, within a quarter, transforms the quality of the company's customer data. The CRM hygiene improvement compounds: the next quarter's renewal forecasts are more accurate, the next quarter's win/loss analysis is meaningful instead of speculative, the next quarter's sales coaching has actual content to point at.

05 — Internal knowledge lookup

The lightest version of the deep knowledge-assistant build. This is "ask anything about our prior work and get a cited answer" applied to whatever knowledge base the firm already has. It pays back more slowly than the first four, but on a 12-month horizon it is consistently among the highest-value workflows we ship — partly because of the obvious ramp-up benefit for new hires, partly because the existence of the system improves the quality of the underlying knowledge base over time.

We put this fifth not because it's least valuable, but because it depends on having a knowledge base worth indexing. If the company's documents are scattered across personal drives, never tagged, and last touched in 2022, the right first move is to fix the source material, not to layer an LLM on top of it. We've watched that mistake play out enough times to be firm about it.

02 · The threeWhat we deliberately leave alone

The list below is harder to write than the one above, because some of the items are workflows that are technically automatable. The reason we don't is that the cost of being wrong is much higher than the cost of leaving the humans in place.

01 — Pricing and discount decisions on real deals

The argument for automating pricing is seductive: the model has access to historical win/loss data, competitor pricing intelligence, and the deal's parameters. It should be able to recommend an optimal price and remove guesswork.

The reason we don't is that the optimal price on most SMB deals is not a function of the data the model can see. It is a function of the buyer's organisational politics, the seller's read of the buyer's urgency, the strategic value of having this logo, the relationship the seller has built with the buyer in the prior 18 months, and a hundred other signals that are in the seller's head and that aren't in any system. We have seen pricing automations recommend numbers that are technically defensible from the data and operationally disastrous, because they're optimising in the wrong space.

We will happily automate quote generation — the document, the line items, the legal terms, the formatting — once the seller has decided the number. We will not automate the number. Use the freed time to give your sales team better win/loss intelligence so the human decision is sharper.

02 — Hiring decisions, especially in the final rounds

There is a reasonable case for automating the very top of the hiring funnel — categorising 800 inbound CVs into "obviously not a fit", "maybe", and "promising" so that humans focus their attention where it matters. We have built this for clients and it can work, with caveats around bias evaluation and an explicit human review of the "obviously not a fit" pile to catch the false negatives.

But automating any later stage of hiring — interview scoring, final-round assessment, offer-versus-no-offer recommendations — is something we don't do and won't pitch. The reasons are stacked: legal exposure under EU and Spanish employment law that is being actively tightened in 2026; the well-documented pattern of AI-driven hiring tools producing confidently wrong assessments based on shallow pattern matching; the harm to the candidate experience of being told that a model rejected them; and the simple fact that hiring is one of the few decisions where being wrong destroys long-term value far more than the time saved.

This is one of the workflows where we are most likely to disagree with another vendor's pitch. Some of them are very good at making this sound responsible. We are not convinced.

03 — Relationship-led customer success outreach

The third one is the most situational. There are SMBs whose customer base is large enough and transactional enough that automated outreach is a perfectly good fit — usage-based reminders, renewal nudges, onboarding sequences. We build those. They work.

But for companies whose business model is built on a small number of high-touch relationships — boutique consultancies, premium B2B services, white-glove account management — automating the outreach that maintains those relationships is a mistake even when it works mechanically. Customers can tell. The reason they're paying a premium is partly the relationship, and the relationship is corroded by interactions that are obviously machine-generated. We have seen this go wrong twice in our own client base, both times with a renewal-nudge automation that produced a measurable bump in nudge-open rates and a measurable drop in renewal sentiment six months later.

The rule we use to draw this line is: if the customer's experience of doing business with the company is a meaningful part of why they pay the price they pay, do not automate the touchpoints that constitute that experience. Automate the back office of those relationships — preparing the human for the conversation, surfacing the right context, tracking the follow-ups — not the conversation itself.

03 · The patternRestated, in a single posture

Automate the parts of the workflow that benefit from speed, consistency, and never sleeping. Leave the humans on the parts that benefit from judgement, context, and discretion.

— The posture, in one sentence

The five-and-three list is more than a recipe; it is a posture. Automate the parts of the workflow that benefit from speed, consistency, and never sleeping. Leave the humans on the parts that benefit from judgement, context, and discretion. The mistake we see most often is companies inverting this — automating the discretion-heavy decision and leaving the humans to do the soul-destroying repetitive intake work.

If your team is currently spending its time on workflows 1 through 5 from our top list, and your founder is spending hers on the three we'd leave alone, you have an opportunity. You also have, almost certainly, a team that's tired, a founder who's overworked, and a competitive position that's quietly weaker than it should be.

If you're not sure where your workflows fall, we publish a free 30-minute audit that maps your current operations against this framework and tells you what we'd start with. We will tell you, candidly, when the answer is "automate nothing for the next six months and fix your data first" — that's about 1 in 5 of the conversations we have, and it's the right answer when it's the right answer.