Madeline, the AI assistant: review fixes

Some people were appearing twice in the end-of-day reports — once under their proper name and once under a name taken from their email address. It split their work across two rows and made team completion read lower than it was.

The fix is written and tested. It needs applying to the live database, which is a two-minute job. Reports filed under the wrong name are merged automatically when it runs.

Someone who replies "stop" to a social message is recorded as opted out immediately and permanently — in that system. The CRM knows nothing about it, so the same person can be messaged again later through a different route.

One shared do-not-contact list that every system checks before sending. This is the one item on the page with a legal edge to it rather than only a quality one.

A chat assistant for the website that answers questions, works out whether an enquiry is a good fit, captures what the sales call needs to know, and books the call — handing straight to a person the moment somebody is frustrated or asks for one.

The assistant's instructions, the qualifying questions, the scoring and the eight follow-up sequences are all written. What remains is assembling them in the CRM, roughly a week, and running it in draft mode for the first few days so every reply is read before it is sent.

A meaningful share of leads land in the CRM with no advert attached to them. They show as "unattributed". That isn't a reporting fault — the information never arrived — but it does mean the advert breakdown is reading from less than the full picture, and budget decisions based on it are weaker than they look.

Worth an hour to trace where the tracking drops off, once the advert figures above are flowing and we can see both ends.

At current volume the reachable audience for a text campaign is a handful of people. A monthly subscription spread across that many contacts costs more per person than sending the messages by hand — and a message written for one person converts better at that size.

Suggested trigger to switch it on: around fifty leads a month, or sooner if replying by hand starts taking longer than five minutes.

Visitors who decline cookies are deliberately not sent to Facebook for advert matching, even though the funnel technically could. UK and EU rules require permission first, and sending it from a server rather than a browser doesn't change that. Those leads still reach the CRM in full — only the advert matching is dropped, which is exactly what the visitor refused.

Madeline can only see the meetings of people who have connected their own account. Until everyone has, "what's on my plate today" is answering from a partial picture, and the team view has gaps that look like nobody is busy.

About five minutes per person on the Integrations page. Nothing to build, and no admin can do it on someone else's behalf — that is deliberate, so one person's inbox is never opened by another.

An invited person has no access until they click the link in their email. Until then they can sign in and see an app with every number at zero, which reads as the product being broken rather than as a seat not yet taken.

Each person clicks their invite link once. Anyone whose invite has expired needs re-inviting.

Our records hold two different surnames for the same person — one from the original team sheet, one from the email address their account was created with. That could be a typo, a name change, or two different people.

Somebody who knows needs to say which is right. The system deliberately refuses to guess: filing one person's work under a colleague's name is a worse outcome than showing it twice.

Anyone connecting their Google account currently sees a warning from Google saying the app has not been verified. Nothing is wrong, and the connection is safe — but it is alarming to read, and some people will stop there rather than click through it.

Two ways to clear it: submit the app to Google for verification, which takes a few weeks, or add each team member as an approved tester, which takes minutes and is enough while the team is this size.

Instagram messages can reach both the advertising follow-up system and the social messaging system. If both are connected to the same account, one message gets two different automated replies in the same conversation. It is the most visible mistake available to us, and a prospect would see it.

Decide which system owns Instagram, write it down, and disconnect the other. Ten minutes to check, and worth doing before any further messaging goes live.

Three systems follow up with leads, and each records progress in its own words. Answering "how many qualified leads did we produce last week" currently means opening two systems and reconciling them by hand.

Either have every system record its leads in the CRM, which suits how the team already works, or keep them separate and report on each. Both are reasonable; the cost of the first is CRM contact volume.

Daily reports filed before the workspace was rebuilt were removed with it. They are not lost permanently — they exist in a database backup — but they are not in the app and will not come back on their own.

Restoring them means standing up a copy of the old database and copying the reports across. Worth deciding whether that history is needed before anyone spends the day on it.

A private page showing how many leads the Facebook ads produced, how many booked a call, and which advert each one came from. It works and is reading live data — but only on the laptop it was built on. Nobody else can open it yet, because it has no home on the internet.

To release it: the hosting account needs permission to create a new project. The account we tried refused — it can view existing projects but not create one. Someone with full access to the hosting account needs to create the project once. After that it updates itself automatically.

This shows what each advert cost, how many people clicked, how far into the videos they watched, and — the useful one — what a booked call actually costs per advert. Facebook can't tell you that on its own, because it doesn't know who booked. The CRM can't either, because it doesn't know what was spent. Putting the two together is the whole point.

To release it: the access key we were given is for a different purpose and can't read advertising figures. A key with advertising-read permission is needed, and the advertising account has to be attached to it. Both are settings changes inside Facebook Business, not development work.

Facebook loses track of a share of conversions to ad blockers and phone privacy settings. Sending them a second time from the server recovers most of it, which makes the adverts optimise against better information. The funnel now sends everything needed for that, including the identifiers that let Facebook match a person without double-counting them.

To finish: generate a conversions key in Facebook Events Manager, paste it into the CRM's Facebook settings, and add the conversion step to the two workflows. Then confirm in Events Manager that each conversion arrives once rather than twice.

Gmail, Outlook and Teams can be connected today, and the Integrations page shows only those three on purpose — a button that cannot do its job reads as a broken product. Slack, Instagram, WhatsApp and Discord are built and tested behind it, switched off until each provider approves the app.

To release them: each provider has to register and approve our app. Slack and Discord are a form and a few minutes. Instagram and WhatsApp go through Meta's review, which takes days to weeks and is outside anyone's control here.

Recommendation is Sendblue, because it connects directly to the CRM, so replies land where the team already works and results can be measured against the advert that produced the lead. The alternative would have needed that connection built from scratch.

Before paying, two things need answering by their sales team: can the fallback to ordinary SMS be switched off (leaving it on pulls you into a registration process you wanted to avoid), and what does outbound messaging actually cost — their public pricing page is ambiguous on that.