The group

Beckford runs pubs with rooms and restaurants across rural Wiltshire, Somerset and Bath. Long-established coaching inns with deep cellars, serious wine lists and kitchens that change with the season. The pilot started at The Talbot Inn in Mells, a working pub with a bar full of the kind of stock that is hard to count and expensive to get wrong.

What was not working

Beckford had been monitoring stock in its own way for years, with its own practices and its own people. That is precisely why the idea of handing the count to an AI landed badly at first.

We were both incredibly sceptical at the beginning that this technology would actually work and do what it said on the tin.

The idea of the AI originally made me quite standoffish, because of how long we have been monitoring our stock in our own practices.

The pilot

Rather than take it on trust, Beckford set a ten-week test with three things to prove. That Peckish would be more accurate than the manual count. That it would be meaningfully faster. That the team on the ground would actually want to use it.

For the first eight weeks, every Peckish count ran alongside a full manual count so the two could be compared item by item.

Setup took one week instead of the two that were planned, and the counts were accurate from the first session. The first fortnight was not frictionless, and Beckford are straightforward about that.

We did hit a few roadblocks in the first couple of weeks. But once we ironed out the creases of narrating too much in the videos, it brought our actual stock count down from an hour and a half to 30 minutes. Once that fell into place, it was a no-brainer for us.

What the counts showed

Across four stocktakes, Peckish averaged 98% accuracy against actuals. The manual count averaged 67%, and never once beat it.

Most of the gap was not small errors. It was whole cases missed. Appletiser, Fanta and San Pellegrino were each counted as zero when there were 48 of each on site. Guinness Zero was counted as five against a true 155. A case of Sauvignon Blanc was counted at 25.5 bottles against 59.5.

By week five the results were consistent enough that the parallel manual count was dropped.

I started to recognise a lot more of the products that the manual count I was working alongside was missing as well. Overall, the retention of our produce would be greater.

What changed

The count came down to 29 minutes, and the argument came down with it. Beckford is now rolling Peckish out across the group, with the bar going first and the kitchen following.

The bigger prize is what accurate, frequent counts unlock. Variance, shrinkage, over-ordering and portion drift all surface while they are still fixable, which is worth considerably more than the counting hours saved.