This is Part 3 of a three-part series on running Claude as my office assistant. Part 1 covered where it’s actually paying off, and Part 2 covered the infrastructure behind it, the backups, the second brain, and the personalised skills. This last part is the honest one: what’s actually gone wrong, the questions people keep asking me, and where I take this next.
Parts 1 and 2 covered what's working: the content pipeline, the CRM, the backups, the second brain, the personalised skills. This last part is the honest one, what's actually gone wrong, the questions people keep asking me, and where I take this next.
I don't want this to read like a sales pitch, so here's what's actually gone wrong.
Claude can't reliably read some technical data straight off a live web page, which matters for SEO work. I built a manual check instead, running it through Google's own testing tool and reading the results back rather than trusting a direct read, which kept part of the audit manual longer than I'd have liked.
Early on, typing directly into CRM cells through the browser corrupted notes, half-overwritten and garbled. That took a proper rebuild with safeguards, so nothing gets edited by accident, only added to, dated and logged. Real engineering, not a prompt tweak.
I also looked at other CRM systems that Claude could integrate with, and had a rebuild on one of them proposed and built out in full detail before it went nowhere, because it was a tool I'd never used and it would have added complexity rather than removed it. The lesson: the right automation fits how you already work, not the flashiest one on the table. Smaller bugs turned up too, like dates picking up an inconsistent format between sheets, the kind you only catch by testing properly rather than trusting it works.
It's also got things wrong from a single misread file name. I'd uploaded a document called Carehome.pdf, background reading for one enquiry, nothing more, and it briefly took the filename as a signal that care homes were a sector I actively work in, rather than one file about one job. Harmless in that instance because I caught it before anything went out, but it's a reminder that this only knows what you tell it, and it will sometimes read more into a stray detail than you meant.
The concern I still carry most is data corruption. I fixed the mid-cell bug, but the underlying fact hasn't changed: my CRM is a Google Sheet, not a proper database, and Claude itself keeps updating in ways I don't control. A change in how it writes, or a new quirk from an update, could damage live data before I notice, and a sheet won't warn you the way a database with real constraints would. And underneath all of it sits the same rule: it's only as good as the information you give it. Every good result here traces back to me having actually written the rate card, the SOPs, the brand voice down properly first. Every bad result traces back to a gap I hadn't filled in yet.
There's a related question I don't have a fully satisfying answer to yet: where the data actually lives. Claude is a US company, and I'm feeding it UK client details, names, emails, sometimes commercially sensitive figures, across email and that spreadsheet. One thing that does help: my sessions are set to private, so nothing I feed in gets used to train the wider model, which means everything generated for GHP+V, the wiki, the CRM setup, the skills, stays mine rather than feeding back into some general pool. That's a real safeguard, but it's not the same thing as a formal GDPR review, and I haven't had one done. If you're weighing this up for your own business, get that checked properly rather than assuming it's fine because everyone else seems to be doing it.
One thing that does reassure me: if Claude ever went offline, or my subscription lapsed, I wouldn't lose the business. The CRM is a Google Sheet and Gmail underneath all of it, both still there and still fully accessible on their own. What I'd lose is the automation, the drafting, the nudges, the sync checks, not the data itself. It's not unlike a colleague calling in sick: the work doesn't run itself that day, but nothing's gone.
None of this is exotic. It's the ordinary friction of getting a new hire up to speed, except this one can also draft your invoice reminders at 9pm.
Two places I want to push it further: deeper marketing automation, so more of the shoot-brief-to-content pipeline runs with less input from me, and financial and admin ops, extending what the Xero and invoicing automation already does. Both build straight off the second brain and the CRM already in place.
There's also a proper enquiry-research step I want to set up, not live yet, still on the list. The idea is a short profile on every new enquiry before I reply: roughly a hundred words on the company, a quick look at their existing photography and video to spot the gaps, and a first read on where GHP+V's offer actually fits what they're missing. Right now that's a lighter, more manual version of the same idea. Building the proper version is next.
If you're starting out, begin with the admin task actually costing you time, not the one that sounds impressive, mine was CRM upkeep and invoice chasing, not content. Keep a draft-and-approve habit for anything client-facing from day one, that one rule has caught more problems than any technical safeguard I've built, though I'm more relaxed about low-stakes, internal-only stuff like CRM housekeeping. Expect a bedding-in period where you're correcting the AI's assumptions about your business about as often as it's saving you time, because that's what builds the second brain that pays off later. Be specific about proof, not adjectives: I don't let anything claim I'm "highly experienced," I say 18 years photographing construction sites across the North West, because that's the sentence a client actually believes. And write down what you already know about your own business first, my rate card, SOPs and named proof points were the real foundation everything else was built on.
I'm still building trust in it, if I'm honest. Nothing client-facing goes out without me reading it first, and I don't expect that to change soon. But it's earning its place, bit by bit, the same way you'd judge any new person in the business: not by what they promise on day one, but by what they've actually got right by the time you stop double-checking everything.
The admin burden of running a creative business alone is real, and I don't think an AI assistant makes it disappear. What it's done is strip out the repetitive scaffolding, so the hours I spend on admin go on the bits that actually need a person: judgement calls, client relationships, and the work itself.
No, and worth saying plainly: this isn't generative AI for pictures. Claude works with text and documents, that's it. Every actual edit still happens in Lightroom and Premiere, by me. If you're picturing something that retouches a shot or generates one from scratch, that's a different category of tool, and not something I've used it for.
Only if I let it. Early drafts came back sounding like nobody in particular, hedged, careful, no actual voice. That's what the brand voice skill fixed (see Part 2): first person, real proof points instead of vague claims, the stock AI phrasing stripped out. It still needs editing every time. It doesn't sound like me without that skill doing the work first.
Roughly broke even so far, once I weigh the time spent building and fixing things against the time it's freed up. I'm on the Pro plan, and the subscription itself is the smallest part of that cost, the real cost was my own hours getting it set up properly. I expect the time saving to tip in my favour from here, since the setup was one-off and the saving isn't, but even on the numbers alone, the peace of mind of knowing the CRM, the follow-ups and the admin are being watched has more than paid for the subscription by itself.
Properly good, but slower than the marketing around it suggests. It's audited the site, dug into Search Console and GA4, rebuilt schema markup, revived blog posts that had gone dormant, and organic sessions have grown off the back of it. But it can't reliably read some of that technical data straight off a live page, so a chunk of the checking still has to happen by hand. A good tool, not a magic one.
Mine wasn't AI-drafted, and that distinction matters. I already had full Terms and Conditions a solicitor had written years ago. What Claude did was turn that into something shorter and in plain language, not invent contract terms from nothing, and I chose to send it on that basis without a fresh solicitor check. That was my call for my situation, not a rule I'd hand anyone else. A rewording pass can still shift meaning without meaning to, so if you're doing the same thing, that's worth weighing up for yourself rather than assuming it's automatically fine because a solicitor was involved somewhere upstream. And if Claude's drafting the substance of a contract with no solicitor behind it at any point, get that checked before it goes near a real client, full stop.
Not for the day-to-day stuff, drafting, CRM updates, content, none of that needed me to write a line of code. The deeper builds in Part 2, the backup hardening and the delivery system, did get technical, but Claude walked me through each step in plain terms rather than expecting me to understand the plumbing myself. Patience gets you further here than coding ability would.
It happened once, properly: typing straight into CRM cells corrupted some notes before I caught it, and fixing it properly took real work, not just an apology. That's exactly why nothing client-facing goes out unsupervised. Low-stakes, internal-only automation runs on its own. Anything a client or the public would see, I read first, every time.
Two months isn't long, and I don't think the tools I'm using now are the tools I'll be using in a year. AI is going to keep changing, whatever's underneath it, whatever it plugs into, whatever it gets better or worse at, and I don't want to be the person who set this up once and then stopped paying attention. I'd rather stay close enough to it that I'm the one deciding how it fits into GHP+V as it evolves, fixing what breaks, dropping what stops earning its place, adopting what's actually useful, than the one who ignored it for a year and found the gap too wide to close. That's the point of writing all three of these posts down honestly rather than as a highlight reel. Not to sell it to you, but to leave myself, and anyone else in the same position, something accurate to build on rather than start over from nothing.
If you’re a fellow photographer, videographer or creative freelancer weighing this up, I’m happy to share more of what’s worked and what hasn’t. Get in touch at office@gregharding.co.uk, or drop a comment below if you’ve been down a similar road. If you’ve landed on this part first, Part 1 covers where it’s paying off and Part 2 covers the setup behind it.