Operations lead, 40-person logistics operator
Read three pricing pages this week, never opened an email
0
score
Revenue operations
AI Sales & Marketing Automation Services
Automate the admin around selling, not the selling. Clean records, a lead score your revenue team can read and argue with, and follow-up that stops the instant a person replies.
ISO 27001 certified
Scoring you can read
Under 3 minutes to first reply
Where the order matters
What CRM automation is
CRM automation removes manual data entry and follow-up from the revenue process by enriching records, scoring leads on observed behaviour, and triggering sequences from pipeline events. The scoring model stays explicit and tunable rather than hidden inside a vendor platform.
153+Projects Delivered
4.9/5Across 153 Reviews
3 MinsAverage Answer Time
100%Code & IP Assigned To You
Four reasons revenue automation gets switched off
Usually by the sales team, quietly, about four months in.
Nobody trusts the score
A number appears on the record and no salesperson can find out why. So they ignore it and work their own list, and the scoring project produced a field nobody reads. A model your team can open, question and adjust beats a better model they will not use.
A sequence kept emailing an existing customer
They were mid-renewal, or had an open support ticket, and the campaign did not know. One of those emails does more reputational damage than the whole sequence generated in pipeline. Suppression rules are not an optimisation, they are the first thing to build.
The data was never good enough to automate on
Three records for the same company, half the industry fields blank, and deal stages that were last accurate in March. Automation applied to that produces wrong routing at speed. Hygiene is unglamorous and it is genuinely the first project.
The outreach read like it was generated
Because it was, at volume, with a token personalisation field. Recipients recognise this immediately now, and the reply rate reflects it. Models are useful for drafting from a real conversation and useless for manufacturing relevance that does not exist.
Open the black box
Build the scoring model yourself
Five anonymised leads with fixed attributes. Move the weights and the ranking reorders. This is roughly the conversation we run with a revenue team, and it is where sales and marketing discover they disagree about what a good lead is.
Size, sector and geography against your ICP.
Opens, visits, content downloaded, events attended.
Pricing page, demo request, procurement questions.
How recently anything happened at all.
Ranked queue
CTO, 400-person manufacturer
Perfect profile, engaged for months, no buying signal yet
0
score
Analyst at a large bank
Opens everything, downloads everything, may be researching
0
score
Founder, 12-person consultancy
Was very active in the spring, silent since
0
score
Procurement manager, healthcare group
Asked about security documentation last Tuesday
0
score
Weight intent heavily and the logistics operator who never opens an email comes first. Weight engagement heavily and the bank analyst does. Both are defensible, and choosing between them is a commercial decision rather than a technical one.
Score is a weighted average of the four attributes, scaled to 100. That is the entire model, and it is deliberately simple enough to argue with in a meeting.
Capabilities
What a marketing automation engagement covers
Ordered the way we build it, which is bottom-up rather than in order of interest.
Data hygiene, first and continuously
Deduplication with a rule for which record survives, enrichment so nobody types a headcount, and a documented answer for which system is authoritative when two disagree. Then it runs on a schedule rather than as a one-off cleanup that decays from the day it finishes.
This is the least popular part of the proposal and the one that determines whether anything above it is true. It runs on the same self-hosted orchestration layer as the rest of your automation.
Suppression before sending
Stop conditions are designed first: a human reply, an open deal, an active support ticket, a frequency cap, a consent change. Enforced at the orchestration layer so a new campaign cannot accidentally bypass them.
Call notes into the CRM
A recording becomes a summary, next steps and updated fields, drafted for the rep to confirm. The highest-value use of a model here, because the alternative is records that were last accurate a month ago.
Learned scoring, later
Once there is enough closed-won and closed-lost history, a model can replace the weights you set by hand. That is predictive analytics territory, and it is a second-year project.
Consent that actually propagates
A preference change updates everywhere, not only where it was captured. Checked at send time rather than at list build. Governed under AI compliance.
Build sequence
How a revenue automation programme runs
Five steps, following the same delivery shape as the rest of the AI and automation practice.
-
Fix the data before anything else
Deduplicate, enrich, and establish which field is authoritative when systems disagree. Nobody asks for this and everything else depends on it. Expect the first weeks of the engagement to look unexciting and to change your reporting more than any dashboard will.
-
Agree what a good lead is, in writing
Sales and marketing usually disagree, and the scoring model is where the disagreement becomes explicit and therefore settleable. This is a workshop rather than a technical task, and skipping it produces a score that one of the two teams quietly ignores.
-
Build scoring you can read
An explicit weighted model held outside the CRM, with weights your revenue team can edit and a visible reason attached to every score. When a rep asks why a lead ranks where it does, the answer should be on the record rather than in a vendor's documentation.
-
Automate follow-up with suppression first
Design the stop conditions before the send conditions. Every sequence we build answers what makes this stop before it answers what makes this send, because the failure mode of the second question is a customer receiving a prospecting email mid-renewal.
-
Instrument and hand over
Dashboards on score distribution, conversion by score band and sequence exit reasons, so you can tell whether the model is right rather than only whether it ran. Then full assignment of definitions, scoring logic, templates and documentation.
Working with the CRM you already have
We do not sell a platform migration as an automation project. Select yours to see what we typically build against it and where the constraints sit.
HubSpot
Strong native automation, which means a good deal of what teams ask us for is already available and unconfigured. We audit what you are paying for before proposing anything external, and frequently the answer is that the gap is process rather than tooling.
Where we do build outside it: scoring logic that needs to be readable and version-controlled, and cross-system suppression that has to consider support tickets HubSpot cannot see.
Usually already possible
Sequences, lifecycle stages, basic scoring
Usually needs building
Readable scoring and cross-system suppression
The parts of a sequence nobody designs
Everyone specifies the emails. These four decide whether the sequence helps or embarrasses you.
A human reply of any kind, an opportunity opening, an active support ticket, a renewal window, an unsubscribe anywhere in the estate. Each is checked immediately before the send rather than when the list was built, because the gap between those two moments is where the embarrassing email happens.
Each campaign looks reasonable in isolation. A contact who qualifies for four of them receives four times as much, and nobody who approved any single campaign approved that. A cap enforced across everything, per person rather than per campaign, is the fix and it belongs in the orchestration layer.
Why did each contact leave: replied, opportunity created, suppressed, unsubscribed, completed without response. Reply rate alone tells you almost nothing; the distribution of exit reasons tells you whether the sequence is working, annoying people, or targeting the wrong list entirely.
When someone replies, a rep needs the thread, the score with its reasoning, and what the contact has actually looked at. Handing over a notification with a name and nothing else recreates the cold-start problem the sequence was meant to solve. Same principle as the escalation design in support automation.
Where AI helps, and where it makes things worse
We are an automation agency saying no to half the AI use cases in our own category. Both lists are honest.
Summarising a call into CRM fields
Real input, verifiable output, and it fixes the record decay that undermines everything else. The rep confirms rather than types.
Drafting a follow-up from actual notes
The relevance already exists in the conversation. The model is doing composition, not inventing a reason to make contact.
Classifying and routing inbound enquiries
Reading a form submission and deciding which team owns it. Bounded, checkable, and it removes a genuine daily delay.
Explaining why a lead scored what it did
Turning the weighted components into a sentence a rep reads in two seconds. Small, and it is what makes the score get used.
What we are usually asked to fix first
Four patterns by sector. Profiles are anonymised.
Consultancies: the CRM is a filing cabinet
Relationships live in partners' heads and the CRM is updated before a review meeting. Call summarisation into records changes this more than any scoring model, because it removes the reason nobody updates it.
Consulting providersManufacturing: enquiry to quote is the bottleneck
The delay is rarely in deciding to quote, it is in gathering specifications and routing to the right engineer. Classification and routing automation returns days per enquiry without touching how anyone sells.
Manufacturing engagementsFinancial services: consent is the whole design
Outreach rules are tighter and the consequences of getting them wrong are regulatory. Suppression and consent propagation get built first and tested hardest, and the scoring work follows once they are provably correct.
Banking and insuranceNon-profit: donor records across three systems
The same supporter appears in the donation platform, the events list and a spreadsheet. Deduplication and a single consent state prevent the appeal that reaches someone twice in a week, which matters more here than anywhere.
Non-profit and NGOThree approaches to lead scoring
The middle column is where almost every mid-market team should be, and it is the one vendors are least interested in selling.
| Criterion | No scoring | Transparent weighted | Machine-learned |
|---|---|---|---|
| Needs historical data | None | None to start | Substantial closed-won and closed-lost |
| Can a rep see why | Not applicable | Yes, component by component | Only with explainability work |
| Adjusting it | Not applicable | Change a weight, see it immediately | Retrain and revalidate |
| Likely accuracy | Whatever the rep judges | Good, once the weights are argued out | Best, given enough clean history |
| Likely adoption | Not applicable | High, because the team helped build it | Low, until it has proved itself |
The third column is genuinely better on accuracy and routinely worse on outcome, because a score nobody acts on has no value regardless of how well it predicts. Move there once the weighted model is being used and its limits are understood. The modelling itself sits under predictive analytics.
What gets scoped alongside this
Neighbouring parts of the AI and automation practice that appear in the same statement of work.
Workflow automation
The orchestration layer where scoring, suppression and enrichment actually run.
AI chatbot development
Qualifying website visitors into scored records rather than into a contact form.
Customer support automation
The other half of the customer record, and the source of half your suppression rules.
Predictive analytics
Where hand-set weights eventually become a learned model, once history supports it.
Generative AI development
The engineering behind call summarisation and drafting, including how it gets evaluated.
The efficiency challenge
The commercial framing, for leaders deciding where an automation budget goes.
Sequencing this against other priorities is what the readiness assessment is for, with platform decisions supported by our IT consulting team. To start, tell us what your CRM looks like or read the help centre.
Clear Answers
CRM and marketing automation questions
The ten that come up in almost every first conversation.
CRM automation removes the manual work around your pipeline: enriching records so nobody types a company size, keeping stages accurate without chasing, scoring leads on observed behaviour, and triggering follow-up from events rather than from someone remembering. The selling stays human.
Most CRMs include automation features and most teams use a fraction of them, usually because the setup work never got done. A large part of these engagements is finishing what you already own, and we will say so rather than proposing new tooling you do not need.
Start with rules you can read. An explicit weighted model that your revenue team can argue with will outperform a black-box score they do not trust, because they will actually act on it. Move to a learned model once you have enough closed-won and closed-lost history to train on honestly.
It can, and most of the damage comes from volume rather than from automation. We design sequences that stop the moment a human replies, cap contact frequency per person, and suppress anyone in an active deal or open support ticket. Automation should make follow-up reliable, not relentless.
It can draft them, and a person should send them. Generated outreach at volume produces content that reads like generated outreach at volume. Where models genuinely help is summarising a call, drafting a follow-up from real notes, and keeping CRM records current without anyone typing.
HubSpot, Salesforce and Zoho most often, and custom systems where a client has built their own. The orchestration layer sits outside the CRM either way, so the same logic survives a platform change instead of being trapped inside vendor workflow builders.
Consent state lives on the record and is checked before every send, not at list-build time. Suppression is enforced at the orchestration layer so it cannot be bypassed by a new campaign, and preference changes propagate across every system rather than only the one where they were captured.
Usually data hygiene, because everything downstream depends on it and nobody enjoys it. Enrichment, deduplication and stage accuracy come before scoring or sequences, since a scoring model built on incomplete records produces confident nonsense.
We execute an SOW and begin discovery within 24 to 48 hours. Data hygiene automation usually reaches production in weeks. Scoring takes longer because it needs history and, more importantly, agreement between sales and marketing on what a good lead actually is.
You do. On final milestone approval, 100% of the workflow definitions, scoring logic, templates and documentation is assigned to you with no recurring license. The scoring weights in particular are kept somewhere your revenue team can edit without a developer.
Tell us what your reps actually do before a call
That list is usually the automation backlog, and it is more useful than any platform audit. We reply in under 3 minutes.